Add files using upload-large-folder tool

data/catalogue/catalogue-gptneo-jsonl-to-meg-ds.slurm

@@ -0,0 +1,50 @@
+#!/bin/bash
+#SBATCH --job-name=catalogue-gptneo-jsonl-to-meg-ds # job name
+#SBATCH --ntasks=1                   # number of MP tasks
+#SBATCH --nodes=1
+#SBATCH --cpus-per-task=40           # number of cores per tasks
+#SBATCH --hint=nomultithread         # we get physical cores not logical
+#SBATCH --time=20:00:00             # maximum execution time (HH:MM:SS)
+#SBATCH --output=logs/catalogue-gptneo-jsonl-to-meg-ds/%x-%j.out          # output file name
+#SBATCH --account=six@cpu
+#SBATCH --array=0-43
+#SBATCH --partition=cpu_p1
+
+set -x -e
+
+source $six_ALL_CCFRWORK/start-prod
+# we need transformers on master as of writing this to get the Neo20b tokenizer
+conda activate teven-tests
+
+# ======= Generate meg-ds file ======
+
+DATASET_PATHS=($(ls -d $six_ALL_CCFRWORK/bigscience-training/jsonls/jsonl_v2/en/*.jsonl))
+DATASET_PATH=${DATASET_PATHS[$SLURM_ARRAY_TASK_ID]}
+
+TOKENIZER_NAME_OR_PATH=EleutherAI/gpt-neox-20b
+
+DATASET_NAME_WITH_JSONL=$(basename $DATASET_PATH)
+DATASET_NAME=${DATASET_NAME_WITH_JSONL:0:-6}
+LANG=$(basename $(dirname $DATASET_PATH))
+SAVE_MEG_DS_DATASET=$six_ALL_CCFRSCRATCH/bigscience-datasets/gptneo-tokenizations/$LANG/"$DATASET_NAME"/meg_ds_"${TOKENIZER_NAME_OR_PATH//\//_}"
+
+mkdir -p $(dirname $SAVE_MEG_DS_DATASET)
+
+if [[ -f "$SAVE_MEG_DS_DATASET"_text_document.bin ]];
+then
+    echo "$SAVE_MEG_DS_DATASET exists."
+    exit 0
+fi
+
+export HF_DATASETS_OFFLINE=1
+export TRANSFORMERS_OFFLINE=1
+
+cd $six_ALL_CCFRWORK/code/Megatron-DeepSpeed
+/usr/bin/time -v python tools/preprocess_data_many_cores.py \
+       --input $DATASET_PATH \
+       --output-prefix $SAVE_MEG_DS_DATASET \
+       --dataset-impl mmap \
+       --tokenizer-type PretrainedFromHF \
+       --tokenizer-name-or-path $TOKENIZER_NAME_OR_PATH \
+       --append-eod \
+       --workers 40

data/catalogue/catalogue-jsonl-to-meg-ds.slurm

@@ -0,0 +1,50 @@
+#!/bin/bash
+#SBATCH --job-name=catalogue-jsonl-to-meg-ds # job name
+#SBATCH --ntasks=1                   # number of MP tasks
+#SBATCH --nodes=1
+#SBATCH --cpus-per-task=40           # number of cores per tasks
+#SBATCH --hint=nomultithread         # we get physical cores not logical
+#SBATCH --time=20:00:00             # maximum execution time (HH:MM:SS)
+#SBATCH --output=logs/catalogue-jsonl-to-meg-ds/%x-%j.out          # output file name
+#SBATCH --account=six@cpu
+#SBATCH --array=0-497
+#SBATCH --partition=cpu_p1
+
+set -x -e
+
+source $six_ALL_CCFRWORK/start-prod
+# We need a specific installation of tokenizers so that it works with bytefallback
+conda activate thomas_data_tooling
+
+# ======= Generate meg-ds file ======
+
+DATASET_PATHS=($(ls -d $six_ALL_CCFRWORK/bigscience-training/jsonls/**/**/*.jsonl))
+DATASET_PATH=${DATASET_PATHS[$SLURM_ARRAY_TASK_ID]}
+
+TOKENIZER_NAME_OR_PATH=bigscience-catalogue-data-dev/byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles
+
+DATASET_NAME_WITH_JSONL=$(basename $DATASET_PATH)
+DATASET_NAME=${DATASET_NAME_WITH_JSONL:0:-6}
+LANG=$(basename $(dirname $DATASET_PATH))
+SAVE_MEG_DS_DATASET=$six_ALL_CCFRSCRATCH/bigscience-datasets/re-tokenizations/$LANG/"$DATASET_NAME"/meg_ds_"${TOKENIZER_NAME_OR_PATH//\//_}"
+
+mkdir -p $(dirname $SAVE_MEG_DS_DATASET)
+
+if [[ -f "$SAVE_MEG_DS_DATASET"_text_document.bin ]];
+then
+    echo "$SAVE_MEG_DS_DATASET exists."
+    exit 0
+fi
+
+export HF_DATASETS_OFFLINE=1
+export TRANSFORMERS_OFFLINE=1
+
+cd $six_ALL_CCFRWORK/code/Megatron-DeepSpeed
+/usr/bin/time -v python tools/preprocess_data_many_cores.py \
+       --input $DATASET_PATH \
+       --output-prefix $SAVE_MEG_DS_DATASET \
+       --dataset-impl mmap \
+       --tokenizer-type PretrainedFromHF \
+       --tokenizer-name-or-path $TOKENIZER_NAME_OR_PATH \
+       --append-eod \
+       --workers 40

data/catalogue/convert_to_jsonl.slurm

@@ -0,0 +1,31 @@
+#!/bin/bash
+#SBATCH --job-name=convert_datasets_to_jsonl # job name
+#SBATCH --ntasks=1                   # number of MP tasks
+#SBATCH --nodes=1
+#SBATCH --cpus-per-task=40           # number of cores per tasks
+#SBATCH --hint=nomultithread         # we get physical cores not logical
+#SBATCH --time=20:00:00             # maximum execution time (HH:MM:SS)
+#SBATCH --output=logs/convert_to_jsonl/%x-%j.out          # output file name
+#SBATCH --array=0-501
+#SBATCH --account=six@cpu
+#SBATCH --partition=cpu_p1
+
+set -x -e
+
+source $six_ALL_CCFRWORK/start-prod
+conda activate thomas_data_tooling
+
+# ======= Generate json file ======
+
+DATASET_PATHS=($(ls -d $six_ALL_CCFRSCRATCH/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/* /gpfsscratch/rech/six/commun/bigscience-datasets/oscar_dedup/*))
+DATASET_PATH=${DATASET_PATHS[$SLURM_ARRAY_TASK_ID]}
+
+BIGSCIENCE_REPO=$WORK/code/big_science/bigscience
+SAVE_JSON_DATASET_PATH_PREFIX=$six_ALL_CCFRSCRATCH/bigscience-datasets/jsonl_v2
+mkdir -p $SAVE_JSON_DATASET_PATH_PREFIX
+
+python $BIGSCIENCE_REPO/data/catalogue/sample_and_convert_to_jsonl.py \
+    --dataset-path $DATASET_PATH\
+    --save-jsonl-dataset-path-prefix $SAVE_JSON_DATASET_PATH_PREFIX \
+    --num-proc 10 \
+    --batch-size 10000

data/catalogue/get_sizes.py

@@ -0,0 +1,58 @@
+import argparse
+import os
+from typing import List, Dict
+from datasets import Dataset, load_dataset
+from multiprocessing import cpu_count
+
+
+def get_size_per_example(texts: List[str]) -> Dict:
+    size_values = [len(text.encode()) for text in texts]
+    examples = {"bytes_len": size_values}
+    return examples
+
+
+def full_size_estimation(
+    ds: Dataset,
+    batch_size: int,
+    content_key: str = "text",
+    num_proc: int = cpu_count(),
+) -> int:
+    if len(ds) == 0:
+        return 0
+
+    ds_with_size = ds.map(
+        get_size_per_example,
+        batched=True,
+        num_proc=num_proc,
+        batch_size=batch_size,
+        input_columns=[content_key],
+        remove_columns=ds.column_names,
+    )
+    len_bytes = sum(ds_with_size["bytes_len"])
+    return len_bytes
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--input-path",
+        type=str,
+        required=True,
+        help="path to jsonl file containing the data",
+    )
+    parser.add_argument(
+        "--output-folder",
+        type=str,
+        required=True,
+        help="path to jsonl file containing the data",
+    )
+    return parser.parse_args()
+
+
+if __name__ == "__main__":
+    args = get_args()
+    ds = load_dataset("json", data_files=args.input_path, split="train")
+    size = full_size_estimation(ds, batch_size=32)
+    dataset_name = os.path.basename(args.input_path)[:-6]
+    with open(os.path.join(args.output_folder, dataset_name), "w") as f:
+        f.write(str(size))

data/catalogue/get_sizes.slurm

@@ -0,0 +1,34 @@
+#!/bin/bash
+#SBATCH --job-name=get_sizes # job name
+#SBATCH --ntasks=1                   # number of MP tasks
+#SBATCH --nodes=1
+#SBATCH --cpus-per-task=40           # number of cores per tasks
+#SBATCH --hint=nomultithread         # we get physical cores not logical
+#SBATCH --time=20:00:00             # maximum execution time (HH:MM:SS)
+#SBATCH --output=logs/get_sizes/%x-%j.out          # output file name
+#SBATCH --account=six@cpu
+#SBATCH --array=0-497
+#SBATCH --partition=cpu_p1
+
+set -x -e
+
+source $six_ALL_CCFRWORK/start-prod
+
+# common repo
+BIGSCIENCE_REPO=/gpfswork/rech/six/commun/code/bigscience
+cd $BIGSCIENCE_REPO
+OUTPUT_FOLDER=$BIGSCIENCE_REPO/sizes_per_dataset/
+mkdir -p $OUTPUT_FOLDER
+
+DATASET_PATHS=($(ls -d $six_ALL_CCFRWORK/bigscience-training/jsonls/**/**/*.jsonl))
+DATASET_PATH=${DATASET_PATHS[$SLURM_ARRAY_TASK_ID]}
+
+DATASET_NAME_WITH_JSONL=$(basename $DATASET_PATH)
+DATASET_NAME=${DATASET_NAME_WITH_JSONL:0:-6}
+
+export HF_DATASETS_OFFLINE=1
+export TRANSFORMERS_OFFLINE=1
+
+/usr/bin/time -v python data/catalogue/get_sizes.py \
+       --input-path $DATASET_PATH \
+       --output-folder $OUTPUT_FOLDER

data/catalogue/load_ratios_meg_ds_format.py

@@ -0,0 +1,54 @@
+import argparse
+import json
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--dataset-ratios-path",
+        type=str,
+        required=True,
+        help="path to JSON file containing input dataset ratios. Values ares dictionary: {'dataset_path': str, 'ratio': float}",
+    )
+    parser.add_argument(
+        "--split",
+        choices=["train", "valid", "test"]
+    )
+    parser.add_argument(
+        "--output-meg-ds-ratio-file",
+        type=str,
+        required=True,
+        help="path to output the language ratio file",
+    )
+    return parser.parse_args()
+
+TOKEN_RANGES={
+    "train": "0:0.950",
+    "valid": "0.950:1.0",
+}
+
+def main():
+    args = get_args()
+
+    token_range = TOKEN_RANGES[args.split]
+
+    with open(args.dataset_ratios_path, "r") as fi:
+        ds_ratios = json.load(fi)
+
+    main_dataset = [f"{ds_ratio['ratio']} {token_range} {ds_ratio['dataset_path']}" for ds_ratio in ds_ratios]
+    if args.split == "train":
+        final_string = f"\"{args.split}: " + ", ".join(main_dataset) + "\"\n"
+    elif args.split == "valid":
+        main_dataset_string = f"\"{args.split}: " + ", ".join(main_dataset) + "\""
+        additional_datasets = [f"\"valid_{ds_ratio['dataset_path'].split('/')[-2]}: 1 {token_range} {ds_ratio['dataset_path']}\"" for ds_ratio in ds_ratios]
+        final_string = main_dataset_string + " " + " ".join(additional_datasets) + "\n"
+    else:
+        raise ValueError(f"unknown split string {args.split}")
+
+
+    # TODO: you can add some extra dataset names for validation/test
+    with open(args.output_meg_ds_ratio_file, "w") as fi:
+        fi.write(final_string)
+
+if __name__ == "__main__":
+    main()

data/catalogue/merge_dataset_per_language.backup.py

@@ -0,0 +1,76 @@
+import argparse
+import json
+from collections import defaultdict
+
+import regex as re
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--dataset-ratios-path",
+        type=str,
+        required=True,
+        help="path to JSON file containing input dataset ratios. Values ares dictionary: {'dataset_path': str, 'ratio': float}",
+    )
+    parser.add_argument(
+        "--split",
+        choices=["train", "valid", "test"]
+    )
+    parser.add_argument(
+        "--meg-ds-dataset-prefix",
+        type=str,
+        required=True,
+        help="We add `lang` to that prefix in order to designate the path for a languages specific dataset."
+    )
+    parser.add_argument(
+        "--output-ratio-file",
+        type=str,
+        required=True,
+        help="path to output the language ratio file",
+    )
+    return parser.parse_args()
+
+
+def main():
+    args = get_args()
+
+    # load training datasets
+    with open(args.dataset_ratios_path, "r") as fi:
+        ds_ratios = json.load(fi)
+
+    # get all individual languages
+    r = re.compile(r"^.*bigscience-catalogue-lm-data/lm_([^_]+)_.*$")
+    datasets_per_language = defaultdict(lambda: [])
+    for ds_ratio in ds_ratios:
+        candidate_lang = r.match(ds_ratio["dataset_path"]).group(1)
+        if candidate_lang == "hi":
+            ds_ratio["lang"] = "indic-hi"
+        else:
+            ds_ratio["lang"] = candidate_lang
+
+        merged_language = ds_ratio["lang"].split("-")[0]
+        # Merge zh languages
+        if candidate_lang in ["zhs", "zht"]:
+            merged_language = "zh"
+
+        datasets_per_language[merged_language].append(ds_ratio)
+
+    # save ratio result into a file (in json format, you can use `load_ratios_meg_ds_format` for get the meg_ds format)
+    language_ds_ratios = [
+        {
+            "ratio": sum([elt["ratio"] for elt in datasets]),
+            "dataset_path": args.meg_ds_dataset_prefix.format(lang=lang),
+            # Additional field to store in case we want to know what's in there.
+            "original_datasets": [
+                dataset["dataset_path"]
+                for dataset in datasets
+            ]
+        }
+        for lang, datasets in datasets_per_language.items()
+    ]
+
+    with open(args.output_ratio_file, "w") as fi:
+        json.dump(language_ds_ratios, fi, indent=2)
+
+if __name__ == "__main__":
+    main()

data/catalogue/merge_dataset_per_languages.backup.slurm

@@ -0,0 +1,60 @@
+#!/bin/bash
+#SBATCH --job-name=catalogue-jsonl-to-meg-ds # job name
+#SBATCH --ntasks=1                   # number of MP tasks
+#SBATCH --nodes=1
+#SBATCH --cpus-per-task=10           # number of cores per tasks
+#SBATCH --hint=nomultithread         # we get physical cores not logical
+#SBATCH --time=20:00:00             # maximum execution time (HH:MM:SS)
+#SBATCH --output=logs/merge-meg-ds/%x-%j.out          # output file name
+#SBATCH --account=six@cpu
+#SBATCH --partition=compil
+
+set -x -e
+
+source $six_ALL_CCFRWORK/start-prod
+# We need a specific installation of tokenizers so that it works with bytefallback
+conda activate thomas_data_tooling
+
+BATCH_ID=0
+TOKENIZER_NAME_OR_PATH=bigscience-catalogue-data-dev/byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v2-dedup-lines-articles
+
+# ======= Generate language ratio file ======
+
+BIGSCIENCE_REPO=$six_ALL_CCFRWORK/code/bigscience
+
+LANGUAGE_RATIOS_PATH=$BIGSCIENCE_REPO/data/catalogue/training_dataset_ratios_batch_${BATCH_ID}_per_language.json
+MEG_DS_DATASET_PREFIX=$six_ALL_CCFRSCRATCH/bigscience-datasets/catalogue/meg-ds-per-lang/{lang}/"${TOKENIZER_NAME_OR_PATH//\//_}"_batch_${BATCH_ID}_text_document
+
+mkdir -p $(dirname $MEG_DS_DATASET_PREFIX)
+
+python $BIGSCIENCE_REPO/data/catalogue/merge_dataset_per_language.py \
+    --dataset-ratios-path $BIGSCIENCE_REPO/data/catalogue/training_dataset_ratios_batch_$BATCH_ID.json \
+    --split train \
+    --meg-ds-dataset-prefix $MEG_DS_DATASET_PREFIX \
+    --output-ratio-file $LANGUAGE_RATIOS_PATH
+
+# ======= Generate merged files ======
+
+MEG_DS_REPO=$six_ALL_CCFRWORK/code/Megatron-DeepSpeed
+pushd $MEG_DS_REPO
+
+readarray -t MERGE_ARGUMENTS < <(python -c "
+import json
+from pathlib import Path
+
+with open(\"$LANGUAGE_RATIOS_PATH\", \"r\") as fi:
+    data = json.load(fi)
+
+for elt in data:
+    Path(elt['dataset_path']).parent.mkdir(parents=True, exist_ok=True)
+
+print('\n'.join([f\"--datasets {' '.join(elt['original_datasets'])} --output-prefix {elt['dataset_path']}\" for elt in data]))
+")
+
+echo $MERGE_ARGUMENTS
+
+for MERGE_ARGUMENT in "${MERGE_ARGUMENTS[@]}"
+do
+	/usr/bin/time -v python -m tools.merge_preprocessed_data \
+        $MERGE_ARGUMENT
+done

data/catalogue/merge_dataset_per_languages.slurm

@@ -0,0 +1,38 @@
+#!/bin/bash
+#SBATCH --job-name=catalogue-jsonl-to-meg-ds # job name
+#SBATCH --ntasks=1                   # number of MP tasks
+#SBATCH --nodes=1
+#SBATCH --cpus-per-task=10           # number of cores per tasks
+#SBATCH --hint=nomultithread         # we get physical cores not logical
+#SBATCH --time=20:00:00             # maximum execution time (HH:MM:SS)
+#SBATCH --output=logs/merge-meg-ds/%x-%j.out          # output file name
+#SBATCH --account=six@cpu
+#SBATCH --array=0-48
+#SBATCH --partition=cpu_p1
+
+set -x -e
+
+source $six_ALL_CCFRWORK/start-prod
+# We need a specific installation of tokenizers so that it works with bytefallback
+conda activate thomas_data_tooling
+
+TOKENIZER_NAME_OR_PATH=bigscience-catalogue-data-dev/byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles
+
+# ======= Generate merged files ======
+
+MEG_DS_REPO=$six_ALL_CCFRWORK/code/Megatron-DeepSpeed
+pushd $MEG_DS_REPO
+
+BASE_PATH=$six_ALL_CCFRSCRATCH/bigscience-datasets/meg-ds_v2
+LANGUAGES=($(ls $BASE_PATH))
+LANG=${LANGUAGES[$SLURM_ARRAY_TASK_ID]}
+
+SAVE_PATH=$six_ALL_CCFRSCRATCH/bigscience-datasets/merged-meg-ds_v2/$LANG/"${TOKENIZER_NAME_OR_PATH//\//_}"_${LANG}_text_document
+# fancy way of collecting all datasets within a folder
+DATASETS=$(ls $six_ALL_CCFRSCRATCH/bigscience-datasets/meg-ds_v2/$LANG/**/*.bin | xargs -I {} python -c "print('{}'[:-4])")
+
+mkdir -p $(dirname $SAVE_PATH)
+
+/usr/bin/time -v python -m tools.merge_preprocessed_data \
+    --datasets $DATASETS \
+    --output-prefix $SAVE_PATH

data/catalogue/merge_dataset_per_languages_v3.slurm

@@ -0,0 +1,38 @@
+#!/bin/bash
+#SBATCH --job-name=catalogue-jsonl-to-meg-ds # job name
+#SBATCH --ntasks=1                   # number of MP tasks
+#SBATCH --nodes=1
+#SBATCH --cpus-per-task=10           # number of cores per tasks
+#SBATCH --hint=nomultithread         # we get physical cores not logical
+#SBATCH --time=20:00:00             # maximum execution time (HH:MM:SS)
+#SBATCH --output=logs/merge-meg-ds/%x-%j.out          # output file name
+#SBATCH --account=six@cpu
+#SBATCH --array=0-48
+#SBATCH --partition=cpu_p1
+
+set -x -e
+
+source $six_ALL_CCFRWORK/start-prod
+# We need a specific installation of tokenizers so that it works with bytefallback
+conda activate thomas_data_tooling
+
+TOKENIZER_NAME_OR_PATH=bigscience-catalogue-data-dev/byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles
+
+# ======= Generate merged files ======
+
+MEG_DS_REPO=$six_ALL_CCFRWORK/code/Megatron-DeepSpeed
+pushd $MEG_DS_REPO
+
+BASE_PATH=$six_ALL_CCFRSCRATCH/bigscience-datasets/meg-ds_v2
+LANGUAGES=($(ls $BASE_PATH))
+LANG=${LANGUAGES[$SLURM_ARRAY_TASK_ID]}
+
+SAVE_PATH=$six_ALL_CCFRSCRATCH/bigscience-datasets/merged-meg-ds_v3_pii/$LANG/"${TOKENIZER_NAME_OR_PATH//\//_}"_${LANG}_text_document
+# fancy way of collecting all datasets within a folder
+DATASETS=$(ls $six_ALL_CCFRSCRATCH/bigscience-datasets/meg-ds_v2/$LANG/**/*.bin | xargs -I {} python -c "print('{}'[:-4])")
+
+mkdir -p $(dirname $SAVE_PATH)
+
+/usr/bin/time -v python -m tools.merge_preprocessed_data \
+    --datasets $DATASETS \
+    --output-prefix $SAVE_PATH

data/catalogue/merge_nigercongo.slurm

@@ -0,0 +1,38 @@
+#!/bin/bash
+#SBATCH --job-name=catalogue-jsonl-to-meg-ds # job name
+#SBATCH --ntasks=1                   # number of MP tasks
+#SBATCH --nodes=1
+#SBATCH --cpus-per-task=10           # number of cores per tasks
+#SBATCH --hint=nomultithread         # we get physical cores not logical
+#SBATCH --time=20:00:00             # maximum execution time (HH:MM:SS)
+#SBATCH --output=logs/merge-meg-ds/%x-%j.out          # output file name
+#SBATCH --account=six@cpu
+#SBATCH --array=0-0
+#SBATCH --partition=cpu_p1
+
+set -x -e
+
+source $six_ALL_CCFRWORK/start-prod
+# We need a specific installation of tokenizers so that it works with bytefallback
+conda activate thomas_data_tooling
+
+TOKENIZER_NAME_OR_PATH=bigscience-catalogue-data-dev/byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles
+
+# ======= Generate merged files ======
+
+MEG_DS_REPO=$six_ALL_CCFRWORK/code/Megatron-DeepSpeed
+pushd $MEG_DS_REPO
+
+BASE_PATH=$six_ALL_CCFRSCRATCH/bigscience-datasets/nigercongo_fusion
+LANGUAGES=($(ls $BASE_PATH))
+LANG=${LANGUAGES[$SLURM_ARRAY_TASK_ID]}
+
+SAVE_PATH=$six_ALL_CCFRSCRATCH/bigscience-datasets/merged-meg-ds_v3_pii/$LANG/"${TOKENIZER_NAME_OR_PATH//\//_}"_${LANG}_text_document
+# fancy way of collecting all datasets within a folder
+DATASETS=$(ls $six_ALL_CCFRSCRATCH/bigscience-datasets/nigercongo_fusion/$LANG/**/*.bin | xargs -I {} python -c "print('{}'[:-4])")
+
+mkdir -p $(dirname $SAVE_PATH)
+
+/usr/bin/time -v python -m tools.merge_preprocessed_data \
+    --datasets $DATASETS \
+    --output-prefix $SAVE_PATH

data/catalogue/oscar-piiv2-jsonl-to-meg-ds.slurm

@@ -0,0 +1,50 @@
+#!/bin/bash
+#SBATCH --job-name=catalogue-jsonl-to-meg-ds # job name
+#SBATCH --ntasks=1                   # number of MP tasks
+#SBATCH --nodes=1
+#SBATCH --cpus-per-task=40           # number of cores per tasks
+#SBATCH --hint=nomultithread         # we get physical cores not logical
+#SBATCH --time=20:00:00             # maximum execution time (HH:MM:SS)
+#SBATCH --output=logs/catalogue-jsonl-to-meg-ds/%x-%j.out          # output file name
+#SBATCH --account=six@cpu
+#SBATCH --array=0-11
+#SBATCH --partition=cpu_p1
+
+set -x -e
+
+source $six_ALL_CCFRWORK/start-prod
+# We need a specific installation of tokenizers so that it works with bytefallback
+conda activate thomas_data_tooling
+
+# ======= Generate meg-ds file ======
+
+DATASET_PATHS=($(ls -d /gpfsscratch/rech/six/commun/bigscience-datasets/pii_no_id_no_num/post/*.jsonl))
+DATASET_PATH=${DATASET_PATHS[$SLURM_ARRAY_TASK_ID]}
+
+TOKENIZER_NAME_OR_PATH=bigscience-catalogue-data-dev/byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles
+
+DATASET_NAME_WITH_JSONL=$(basename $DATASET_PATH)
+DATASET_NAME=${DATASET_NAME_WITH_JSONL:0:-6}
+LANG=$(basename $(dirname $DATASET_PATH))
+SAVE_MEG_DS_DATASET=$six_ALL_CCFRSCRATCH/bigscience-datasets/oscar_pii_v2/$LANG/"$DATASET_NAME"/meg_ds_"${TOKENIZER_NAME_OR_PATH//\//_}"
+
+mkdir -p $(dirname $SAVE_MEG_DS_DATASET)
+
+if [[ -f "$SAVE_MEG_DS_DATASET"_text_document.bin ]];
+then
+    echo "$SAVE_MEG_DS_DATASET exists."
+    exit 0
+fi
+
+export HF_DATASETS_OFFLINE=1
+export TRANSFORMERS_OFFLINE=1
+
+cd $six_ALL_CCFRWORK/code/Megatron-DeepSpeed
+/usr/bin/time -v python tools/preprocess_data_many_cores.py \
+       --input $DATASET_PATH \
+       --output-prefix $SAVE_MEG_DS_DATASET \
+       --dataset-impl mmap \
+       --tokenizer-type PretrainedFromHF \
+       --tokenizer-name-or-path $TOKENIZER_NAME_OR_PATH \
+       --append-eod \
+       --workers 40

data/catalogue/sample_and_convert_to_jsonl.py

@@ -0,0 +1,650 @@
+import argparse
+import logging
+import re
+from pathlib import Path
+
+from datasets import Dataset, load_from_disk
+from datasets.utils.logging import set_verbosity_info
+from numpy.random import default_rng
+
+set_verbosity_info()
+logger = logging.getLogger(__name__)
+rng = default_rng(42)
+
+CATALOGUE_DATASETS = {
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_ca_enriched_conllu_ancora_for_ml_training": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_ca_parlament_parla": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-pa_ted_talks_iwslt": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_en_odiencorp": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-as_ted_talks_iwslt": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-as_wiktionary_filtered": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_fr_book_dash_books": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_vi_uit_vsmec": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-ur_mkb": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_id_indo4b_talpco": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_en_book_dash_books": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_vi_vietnamese_students_feedback": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_ca_xquad_ca": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-hi_wikimedia_filtered": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-mr_opus100": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_pseudocrawl-filtered_401_www_elperiodicodemexico_com": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_id_indonesian_frog_storytelling_corpus": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-pa_wikibooks_filtered": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_ca_opus100": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-or_mkb": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-te_opus100": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_nigercongo-tum_aggregated": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_vi_opus100": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-kn_opus100": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-pa_opus100": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-kn_ted_talks_iwslt": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_id_indonli": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_nigercongo-bm_aggregated": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_nigercongo-ki_aggregated": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_eu_ted_talks_iwslt": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-te_mkb": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-hi_mkb": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-gu_ted_talks_iwslt": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-ur_wikiquote_filtered": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-mr_wikiquote_filtered": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_ca_viquiquad": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-bn_mkb": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_eu_wikiquote_filtered": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-gu_mkb": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-mr_mkb": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-te_ted_talks_iwslt": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_nigercongo-ak_aggregated": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_nigercongo-ts_aggregated": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_nigercongo-st_aggregated": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-ml_ted_talks_iwslt": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-ml_mkb": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_ca_vilaquad": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-or_opus100": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-gu_opus100": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-bn_bangla_sentiment_classification_datasets": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-gu_wiktionary_filtered": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-hi_wikiversity_filtered": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_nigercongo-ny_aggregated": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_pt_wikimedia_filtered": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_nigercongo-tw_aggregated": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-kn_wikiquote_filtered": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-ta_mkb": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-mr_wiktionary_filtered": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-ur_ted_talks_iwslt": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-gu_wikiquote_filtered": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_nigercongo-tn_aggregated": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-ur_open_subtitles": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-te_wikiquote_filtered": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_id_wikiquote_filtered": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-ur_wiktionary_filtered": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_ca_wikimedia_filtered": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-ml_wikibooks_filtered": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_nigercongo-ln_aggregated": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-hi_wikivoyage_filtered": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-bn_ted_talks_iwslt": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_id_opus100": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-te_open_subtitles": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-bn_wikibooks_filtered": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-te_wikibooks_filtered": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_nigercongo-nso_aggregated": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_vi_vinbigdata_asr_vlsp_2020": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-mr_ted_talks_iwslt": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-ta_ted_talks_iwslt": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-mr_wikibooks_filtered": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_id_wikimedia_filtered": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_en_scielo": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-ta_open_subtitles": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-ml_wikiquote_filtered": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_nigercongo-rn_aggregated": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_vi_vinbigdata_mt_vlsp_2020": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_id_indo4b_parallel": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_id_indo4b_bppt": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_ca_wiktionary_filtered": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_nigercongo-wo_aggregated": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_ar_wikiversity_filtered": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_ca_ted_talks_iwslt": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-ta_opus100": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_vi_wikiquote_filtered": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_zht_qedcorpus": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-ta_wikibooks_filtered": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-hi_qedcorpus": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-ur_wikibooks_filtered": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-as_opus100": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_nigercongo-lg_aggregated": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_en_pseudocrawl-filtered_502_www_ricemedia_co": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_zhs_qedcorpus": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-hi_opus100": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-hi_wikiquote_filtered": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-bn_wiktionary_filtered": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-hi_ted_talks_iwslt": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_nigercongo-fon_aggregated": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_id_recibrew": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-ur_opus100": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-bn_wikivoyage_filtered": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_ar_wiktionary_filtered": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-ta_wikiquote_filtered": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_ca_wikiquote_filtered": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_nigercongo-sn_aggregated": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-kn_wiktionary_filtered": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_ar_arabench": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_fr_qedcorpus": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_pt_opus100": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_pseudocrawl-filtered_30_www_radiocable_com": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-hi_open_subtitles": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_eu_wiktionary_filtered": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-ml_opus100": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_vi_wikivoyage_filtered": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_ca_wikibooks_filtered": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_ca_wikinews_filtered": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_ar_qedcorpus": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_nigercongo-zu_aggregated": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_pt_wikivoyage_filtered": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_ar_wikiquote_filtered": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_wiktionary_filtered": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-as_samanantar": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_eu_wikibooks_filtered": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_id_ted_talks_iwslt": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_zh_wikiquote_filtered": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_zh_pseudocrawl-filtered_674_ai_baidu_com": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_ar_wikibooks_filtered": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_fr_ester": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-ml_wiktionary_filtered": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_zh_wikiversity_filtered": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-pa_wiktionary_filtered": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_nigercongo-ig_aggregated": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_pt_qedcorpus": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_zh_wikinews_filtered": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_zh_wikivoyage_filtered": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_id_wikibooks_filtered": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_en_pseudocrawl-filtered_492_www_vivawoman_net": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_wikiquote_filtered": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-ml_pib": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-te_pib": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_vi_wikibooks_filtered": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-ta_wikinews_filtered": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-gu_pib": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_wikiversity_filtered": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-bn_bengali_question_answering": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_pseudocrawl-filtered_244_www_df_cl": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_ca_open_subtitles": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_pt_wikiquote_filtered": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_scielo": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_zh_ted_talks_iwslt": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_en_pseudocrawl-filtered_485_blog_moneysmart_sg": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_vi_ted_talks_iwslt": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_en_pseudocrawl-filtered_488_dailyvanity_sg": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_en_pseudocrawl-filtered_696_www_oercommons_org": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_vi_wiktionary_filtered": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_wikinews_filtered": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_ted_talks_iwslt": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_en_pseudocrawl-filtered_483_alvinology_com": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_id_indo4b_kompas": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_qedcorpus": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_nigercongo-xh_aggregated": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-or_odiencorp": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_en_ted_talks_iwslt": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_ar_opus100": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_eu_open_subtitles": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_ar_habibi": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_pt_wikiversity_filtered": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_zh_wikibooks_filtered": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-bn_pib": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-pa_pib": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-bn_opus100": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_fr_wikiquote_filtered": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_fr_ted_talks_iwslt": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_eu_wikisource_filtered": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_pt_pseudocrawl-filtered_672_pt_globalvoices_org": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_ar_labr": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_en_pseudocrawl-filtered_339_www_actasanitaria_com": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-te_wiktionary_filtered": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_ar_ted_talks_iwslt": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-mr_pib": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_ar_wikinews_filtered": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-or_pib": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-bn_open_subtitles": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_fr_wikinews_filtered": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_fr_wikivoyage_filtered": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-ur_pib": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_wikivoyage_filtered": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-hi_wikibooks_filtered": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_eu_pseudocrawl-filtered_563_ahotsak_eus": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_id_indo4b_tempo": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_pseudocrawl-filtered_32_www_elexpresso_com": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_id_indosum": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-ml_open_subtitles": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_pseudocrawl-filtered_91_www_diario26_com": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_nigercongo-rw_aggregated": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_pt_ted_talks_iwslt": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_eu_opus100": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-hi_wiktionary_filtered": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_en_pseudocrawl-filtered_500_www_asiaone_com_singapore": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_fr_pseudocrawl-filtered_530_www_mediapart_fr": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-ta_pib": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_pt_wikibooks_filtered": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_en_pseudocrawl-filtered_470_forums_hardwarezone_com_sg": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_pseudocrawl-filtered_223_www_eltambor_es": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_id_indo4b_jw300": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_pseudocrawl-filtered_219_www_aguasresiduales_info": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_pseudocrawl-filtered_376_www_elpopular_com_ar": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_pseudocrawl-filtered_62_www_lapagina_com_sv": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-or_wikisource_filtered": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_vi_data_on_covid_19_news_coverage_in_vietnam": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_en_pseudocrawl-filtered_548_remezcla_com": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_pt_wiktionary_filtered": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_en_pseudocrawl-filtered_499_www_today_com_news": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_en_wikinews_filtered": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_vi_vntq_corpus_big": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_id_pseudocrawl-filtered_545_www_detik_com": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-or_wiktionary_filtered": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-or_wikipedia": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_eu_pseudocrawl-filtered_637_www_argia_eus": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_en_pseudocrawl-filtered_501_theindependent_sg": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_pseudocrawl-filtered_430_www_eldiario_ec": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_pseudocrawl-filtered_420_www_retema_es": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-as_wikipedia": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_ar_kalimat": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_en_pseudocrawl-filtered_487_thesmartlocal_com": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_ca_wikisource_filtered": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_pseudocrawl-filtered_153_financialfood_es": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_pt_wikinews_filtered": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_id_wikisource_filtered": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_wikibooks_filtered": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-gu_samanantar": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_pseudocrawl-filtered_158_www_diariodeleon_es": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_fr_pseudocrawl-filtered_599_fr_globalvoices_org": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_pseudocrawl-filtered_392_www_muypymes_com": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_eu_pseudocrawl-filtered_506_goiena_eus": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_fr_wikibooks_filtered": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-hi_pib": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_pseudocrawl-filtered_231_ojo_pe": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_pseudocrawl-filtered_167_www_ambientum_com": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_id_wiktionary_filtered": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_vi_vietnamese_poetry": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_pseudocrawl-filtered_118_www_elheraldo_hn": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_pseudocrawl-filtered_233_www_dinero_com": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_eu_pseudocrawl-filtered_635_www_berria_eus": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_en_pseudocrawl-filtered_395_www_evwind_es": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_pseudocrawl-filtered_182_correodelsur_com": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_fr_wikiversity_filtered": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_pseudocrawl-filtered_56_www_eluniverso_com": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_pt_project_gutenberg": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_pseudocrawl-filtered_250_www_cooperativa_cl": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_pseudocrawl-filtered_374_www_talcualdigital_com": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_nigercongo-yo_aggregated": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_pseudocrawl-filtered_315_lasillavacia_com": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_pseudocrawl-filtered_207_elimpulso_com": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_pseudocrawl-filtered_276_radio_uchile_cl": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_pseudocrawl-filtered_325_www_laprensa_hn": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_pseudocrawl-filtered_294_www_laopinion_com_co": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-gu_wikipedia": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_pseudocrawl-filtered_34_www_losandes_com_ar": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_pseudocrawl-filtered_90_peru_com": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_pseudocrawl-filtered_23_www_elconfidencialdigital_com": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_pseudocrawl-filtered_125_www_noticiasde_es": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_pseudocrawl-filtered_359_www_efeverde_com": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-mr_samanantar": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-pa_wikisource_filtered": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_en_qedcorpus": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_pseudocrawl-filtered_381_www_cuartopoder_es": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_pseudocrawl-filtered_518_www_elcolombiano_com": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_pseudocrawl-filtered_280_salamancartvaldia_es": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_pseudocrawl-filtered_341_es_cointelegraph_com": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-or_samanantar": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_pseudocrawl-filtered_165_www_ticbeat_com": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-te_samanantar": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_pseudocrawl-filtered_641_es_globalvoices_org": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_vi_wikisource_filtered": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_pseudocrawl-filtered_209_misionesonline_net": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-pa_wikipedia": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-ta_wiktionary_filtered": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_pseudocrawl-filtered_246_www_eldiarionuevodia_com_ar": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_pseudocrawl-filtered_86_www_motorpasion_com": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_pseudocrawl-filtered_257_www_diaridetarragona_com": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-kn_samanantar": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_pseudocrawl-filtered_53_www_expreso_ec": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_pseudocrawl-filtered_367_elcorreoweb_es": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_en_pseudocrawl-filtered_159_www_postcrescent_com": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_pseudocrawl-filtered_422_www_formulatv_com": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_zh_project_gutenberg": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_zh_pseudocrawl-filtered_503_www_zaobao_com_sg": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_en_pseudocrawl-filtered_638_globalvoices_org": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_pseudocrawl-filtered_78_www_listindiario_com": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_pseudocrawl-filtered_277_www_entornointeligente_com": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_pseudocrawl-filtered_404_www_telam_com_ar": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_pseudocrawl-filtered_431_www_elperiodicoextremadura_com": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_pseudocrawl-filtered_254_diario_mx": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_pseudocrawl-filtered_169_www_el_carabobeno_com": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_pseudocrawl-filtered_417_www_radiolaprimerisima_com": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_pseudocrawl-filtered_21_www_elperiodicodearagon_com": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_vi_open_subtitles": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_en_pseudocrawl-filtered_498_www_channelnewsasia_com": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_en_pseudocrawl-filtered_304_www_semana_com": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_pseudocrawl-filtered_386_www_prensalibre_com": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-pa_samanantar": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-ur_urdu-monolingual-corpus": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_pseudocrawl-filtered_286_www_nacion_com": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_pseudocrawl-filtered_67_www_elpais_cr": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_pseudocrawl-filtered_405_www_emol_com": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_pseudocrawl-filtered_317_diariocorreo_pe": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-as_wikisource_filtered": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-gu_wikisource_filtered": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-mr_wikipedia": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_id_pseudocrawl-filtered_512_kumparan_com": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_pseudocrawl-filtered_237_www_cronista_com": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_pseudocrawl-filtered_116_www_latribuna_hn": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_id_indonesian_news_corpus": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_pseudocrawl-filtered_172_www_rionegro_com_ar": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_pseudocrawl-filtered_226_www_ole_com_ar": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_en_wikiversity_filtered": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_pseudocrawl-filtered_287_www_cibercuba_com": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_fr_wiktionary_filtered": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_id_pseudocrawl-filtered_572_tirto_id": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-bn_samanantar": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_pseudocrawl-filtered_157_www_elsoldemexico_com_mx": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_pseudocrawl-filtered_44_ladiaria_com_uy": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_ca_catalan_government_crawling": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_en_wikivoyage_filtered": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_pseudocrawl-filtered_213_www_hola_com": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_pseudocrawl-filtered_324_gestion_pe": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_pseudocrawl-filtered_28_www_fayerwayer_com": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_nigercongo-sw_aggregated": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-ml_samanantar": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_pseudocrawl-filtered_354_www_lagaceta_com_ar": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_pt_wikisource_filtered": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_en_pseudocrawl-filtered_534_www_nairaland_com": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-ta_samanantar": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_pseudocrawl-filtered_253_www_debate_com_mx": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_pseudocrawl-filtered_248_www_telesurtv_net": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_pseudocrawl-filtered_406_www_americaeconomia_com": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_pseudocrawl-filtered_130_www_elperiodicomediterraneo_com": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_pseudocrawl-filtered_232_tn_com_ar": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-ur_leipzig_wortschatz_urdu_newscrawl_2016_sentences": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_pseudocrawl-filtered_71_www_rtve_es": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_pseudocrawl-filtered_220_www_vanguardia_com_mx": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_id_pseudocrawl-filtered_549_www_cnnindonesia_com": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_project_gutenberg": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-ur_wikipedia": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_vi_vietai_sat": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_ca_tecla": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_pseudocrawl-filtered_203_www_que_es": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_pseudocrawl-filtered_320_www_paginasiete_bo": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_pseudocrawl-filtered_181_noticiassin_com": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_id_wikipedia": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-mr_wikisource_filtered": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_en_the_pile_europarl": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_pseudocrawl-filtered_675_www_elespectador_com": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_pseudocrawl-filtered_146_www_perfil_com": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-ml_wikisource_filtered": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_the_pile_europarl": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_pt_the_pile_europarl": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_ar_brad_2": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_pseudocrawl-filtered_79_www_laopiniondemurcia_es": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_pseudocrawl-filtered_229_www_expansion_com": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-kn_wikipedia": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_pseudocrawl-filtered_299_www_lne_es": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_pseudocrawl-filtered_263_www_lasexta_com": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_fr_the_pile_europarl": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_en_wikiquote_filtered": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_pseudocrawl-filtered_136_valenciaplaza_com": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_pseudocrawl-filtered_256_www_laprovincia_es": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_pseudocrawl-filtered_373_www_farodevigo_es": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_en_royal_society_corpus": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_pseudocrawl-filtered_245_www_noticiasdenavarra_com": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-hi_pseudocrawl-filtered_515_www_aajtak_in": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-kn_wikisource_filtered": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_en_wiktionary_filtered": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-ml_wikipedia": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_pseudocrawl-filtered_58_www_levante_emv_com": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_pseudocrawl-filtered_198_www_eleconomista_es": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_pseudocrawl-filtered_249_www_telecinco_es": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_pseudocrawl-filtered_288_www_marca_com": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_ar_ksucca": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_pseudocrawl-filtered_189_www_eleconomista_com_mx": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_pt_scielo": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_pseudocrawl-filtered_103_www_elmostrador_cl": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_vi_wikipedia": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_eu_wikipedia": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-ur_leipzig_wortschatz_urdu-pk_web_2019_sentences": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_pseudocrawl-filtered_211_www_elcomercio_com": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-te_wikisource_filtered": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-bn_indic_nlp_corpus": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_ar_sanad": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_en_wikibooks_filtered": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_pseudocrawl-filtered_429_cadenaser_com": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-hi_wikipedia": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_id_open_subtitles": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-te_wikipedia": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_pseudocrawl-filtered_267_www_elperiodico_com_es": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-hi_wikisource_filtered": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_ar_pseudocrawl-filtered_595_mawdoo3_com": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_pseudocrawl-filtered_215_www_lainformacion_com": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-ta_wikipedia": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-bn_wikipedia": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_zh-tw_wikipedia": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_zh-cn_wikipedia": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_pseudocrawl-filtered_255_elcomercio_pe": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-gu_indic_nlp_corpus": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-or_indic_nlp_corpus": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_wikisource_filtered": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_id_indonesian_news_articles_2017": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_pseudocrawl-filtered_409_www_proceso_com_mx": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_zh_open_subtitles": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_en_pseudocrawl-filtered_510_timesofindia_indiatimes_com": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_pseudocrawl-filtered_349_www_eltiempo_com": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-hi_samanantar": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_pseudocrawl-filtered_424_www_lavanguardia_com": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_pseudocrawl-filtered_100_www_aporrea_org": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_vi_vinbigdata_monolingual_vlsp_2020": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_eu_bsbasque": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_en_pseudocrawl-filtered_497_www_straitstimes_com": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_pseudocrawl-filtered_396_www_eldiario_es": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_pt_wikipedia": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-pa_indic_nlp_corpus": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_ar_tashkeela": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-mr_indic_nlp_corpus": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_zh_du_reader": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_ar_wikisource_filtered": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_pseudocrawl-filtered_20_www_clarin_com": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_en_pseudocrawl-filtered_689_www_abc_net_au": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-hi_pseudocrawl-filtered_667_www_bhaskar_com": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_pseudocrawl-filtered_63_www_lanacion_com_ar": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-kn_indic_nlp_corpus": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_pseudocrawl-filtered_333_www_elmundo_es": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_fr_project_gutenberg": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_fr_pseudocrawl-filtered_550_www_lemonde_fr": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_zh_multi_un_2": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-te_indic_nlp_corpus": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_ca_wikipedia": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-ta_wikisource_filtered": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_zh_uncorpus": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_multi_un_2": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_ca_catalan_general_crawling": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_wikipedia": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_en_multi_un_2": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-ne_unsupervised_cross_lingual_representation_learning_at_scale": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-ml_indic_nlp_corpus": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_fr_multi_un_2": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_ar_multi_un_2": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_ar_wikipedia": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-bn_wikisource_filtered": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_pt_open_subtitles": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_fr_wikipedia": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_fr_open_subtitles": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_ar_open_subtitles": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-hi_iitb_english_hindi_corpus": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_en_uncorpus": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_uncorpus": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_en_no_code_stackexchange": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_ar_uncorpus": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-ta_indic_nlp_corpus": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_fr_uncorpus": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_es_open_subtitles": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-hi_indic_nlp_corpus": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_vi_binhvq_news_corpus": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_ca_catalan_textual_corpus": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_indic-bn_bangla_lm": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_en_wikipedia": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_en_open_subtitles": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_fr_wikisource_filtered": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_ar_openiti_proc": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_ar_arabic_billion_words": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_pt_brwac": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_en_project_gutenberg": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_en_the_pile_uspto": 0.5441176470588235, # ((350 [expected] - ( 326 [catalogue_en] - 251 [s2orc] - 21 [uspto]) ) * 1/2 [catalogue_en_proportion]) / (251 [s2orc] + 21 [uspto])
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_code_stackexchange": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_fr_hal_archives_ouvertes": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_code_github-no-gpl": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_zh_wudaocorpora": 1.,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_en_s2orc_ai2_pdf_parses": 0.5441176470588235, # ((350 [expected] - ( 326 [catalogue_en] - 251 [s2orc] - 21 [uspto]) ) * 1/2 [catalogue_en_proportion]) / (251 [s2orc] + 21 [uspto])
+}
+OSCAR_DATASETS = {
+    # oscar
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/oscar_dedup/ar": 1,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/oscar_dedup/bn": 1,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/oscar_dedup/ca": 1,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/oscar_dedup/en": 0.13454545454545455, # ((350 [expected] - ( 326 [catalogue_en] - 251 [s2orc] - 21 [uspto]) ) * 1/2 [oscar_en proportion] ) / 1_100 [oscar_en]
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/oscar_dedup/es": 1,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/oscar_dedup/eu": 1,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/oscar_dedup/fr": 1,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/oscar_dedup/hi": 1,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/oscar_dedup/id": 1,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/oscar_dedup/pt": 1,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/oscar_dedup/ur": 1,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/oscar_dedup/vi": 1,
+    "/gpfsscratch/rech/six/commun/bigscience-datasets/oscar_dedup/zh": 1
+}
+assert set(OSCAR_DATASETS.keys()).isdisjoint(set(CATALOGUE_DATASETS.keys()))
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--dataset-path", choices=list(set(CATALOGUE_DATASETS.keys()) | set(OSCAR_DATASETS.keys())), type=str, required=True,
+        help="Dataset path."
+    )
+    parser.add_argument(
+        "--save-jsonl-dataset-path-prefix", type=Path, required=True,
+        help="Where to output json file. Files will be save in `{args.save_jsonl_dataset_path_prefix}/{lang}/{dataset_name}"
+    )
+    parser.add_argument(
+        "--num-proc", type=int, default=1
+    )
+    parser.add_argument(
+        "--batch-size", type=int
+    )
+    return parser.parse_args()
+
+
+catalogue_language_regex = re.compile(
+    r"^/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/lm_([^_]+)_.*$"
+)
+normalise_catalogue_dataset_name_regex = re.compile(
+    r"^/gpfsscratch/rech/six/commun/bigscience-datasets/catalogue/clean_v2/bigscience-catalogue-lm-data/(.*)$"
+)
+def get_catalogue_language(dataset_name: str) -> str:
+    lang_candidate = catalogue_language_regex.match(dataset_name).group(1)
+
+    # Normalise chinese languages, so that we only consider simplified and traditional chinese as the two chinese languages
+    if lang_candidate in ["zh", "zhs", "zh-cn"]:
+        lang_candidate = "zhs"
+    elif lang_candidate in ["zht", "zh-tw"]:
+        lang_candidate = "zht"
+    else:
+        assert lang_candidate[:2] != "zh"
+
+    return lang_candidate
+
+oscar_to_bs_language = {
+    "ar": "ar",
+    "bn": "indic-bn",
+    "ca": "ca",
+    "en": "en",
+    "es": "es",
+    "eu": "eu",
+    "fr": "fr",
+    "hi": "indic-hi",
+    "id": "id",
+    "pt": "pt",
+    "ur": "indic-ur",
+    "vi": "vi",
+    "zh": "zhs"
+}
+oscar_language_regex = re.compile(
+    r"^/gpfsscratch/rech/six/commun/bigscience-datasets/oscar_dedup/(.*)$"
+)
+def get_oscar_language(dataset_name: str) -> str:
+    return oscar_to_bs_language[oscar_language_regex.match(dataset_name).group(1)]
+
+
+def sample_dataset(dataset: Dataset, ratio: float) -> Dataset:
+    logger.info(f"Ratio: {ratio}")
+    if ratio >= 1:
+        return dataset
+
+    num_samples = int(len(dataset) * ratio)
+    indices = rng.choice(len(dataset), size=num_samples, replace=False, shuffle=False)
+    return dataset.select(indices)
+
+def main():
+    logging.basicConfig(
+        format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
+        datefmt="%m/%d/%Y %H:%M:%S",
+        level=logging.INFO,
+    )
+    args = get_args()
+    logger.info(f"** The job is runned with the following arguments: **\n{args}\n **** ")
+
+    # Compute save path
+    save_path: Path
+    if args.dataset_path in CATALOGUE_DATASETS:
+        lang = get_catalogue_language(args.dataset_path)
+        filename = f"{normalise_catalogue_dataset_name_regex.match(args.dataset_path).group(1)}.jsonl"
+        save_path = Path(args.save_jsonl_dataset_path_prefix) / lang / filename
+    elif args.dataset_path in OSCAR_DATASETS:
+        lang = get_oscar_language(args.dataset_path)
+        save_path = Path(args.save_jsonl_dataset_path_prefix) / lang / f"lm_{lang}_oscar.jsonl"
+    else:
+        raise NotImplementedError
+
+    # Saved dataset don't require us to re-run de pipeline
+    if save_path.exists():
+        logger.info(f"{save_path} already exists. Exiting early.")
+        return
+
+    # load_dataset
+    logger.info(f"Loading {args.dataset_path}")
+    if args.dataset_path in CATALOGUE_DATASETS:
+        ds = load_from_disk(Path(args.dataset_path) / "final")
+    elif args.dataset_path in OSCAR_DATASETS:
+        ds = load_from_disk(args.dataset_path)
+    else:
+        raise NotImplementedError
+
+    # remove all columns except text
+    logger.info(f"Removing all columns except `text`")
+    columns_to_remove = set(ds.column_names)
+    columns_to_remove.remove("text")
+    ds = ds.remove_columns(list(columns_to_remove))
+
+    # sample dataset according to ratio
+    logger.info(f"Sampling dataset according to given ratio")
+    if args.dataset_path in CATALOGUE_DATASETS:
+        ds = sample_dataset(ds, CATALOGUE_DATASETS[args.dataset_path])
+    elif args.dataset_path in OSCAR_DATASETS:
+        ds = sample_dataset(ds, OSCAR_DATASETS[args.dataset_path])
+    else:
+        raise NotImplementedError
+
+    # save to json
+    logger.info(f"Saving to {save_path}")
+    tmp_save_path = Path(save_path.parent, f"tmp-{save_path.name}")
+    tmp_save_path.parent.mkdir(parents=True, exist_ok=True)
+    ds.to_json(
+        tmp_save_path,
+        num_proc=args.num_proc,
+        batch_size=args.batch_size
+    )
+    tmp_save_path.rename(save_path)
+
+if __name__ == "__main__":
+    main()

data/catalogue/training_dataset_ratios.json

@@ -0,0 +1,198 @@
+[
+    {
+    "dataset_path": "/gpfswork/rech/six/commun/bigscience-training/merged-meg-ds_v2/nigercongo-tum/bigscience-catalogue-data-dev_byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles_nigercongo-tum_text_document",
+    "ratio": 1.6908712906621913e-07
+    },
+    {
+    "dataset_path": "/gpfswork/rech/six/commun/bigscience-training/merged-meg-ds_v2/nigercongo-ki/bigscience-catalogue-data-dev_byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles_nigercongo-ki_text_document",
+    "ratio": 3.825549510416119e-07
+    },
+    {
+    "dataset_path": "/gpfswork/rech/six/commun/bigscience-training/merged-meg-ds_v2/nigercongo-bm/bigscience-catalogue-data-dev_byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles_nigercongo-bm_text_document",
+    "ratio": 4.096353790340665e-07
+    },
+    {
+    "dataset_path": "/gpfswork/rech/six/commun/bigscience-training/merged-meg-ds_v2/nigercongo-ak/bigscience-catalogue-data-dev_byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles_nigercongo-ak_text_document",
+    "ratio": 6.938762268345196e-07
+    },
+    {
+    "dataset_path": "/gpfswork/rech/six/commun/bigscience-training/merged-meg-ds_v2/nigercongo-ts/bigscience-catalogue-data-dev_byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles_nigercongo-ts_text_document",
+    "ratio": 7.058099259657895e-07
+    },
+    {
+    "dataset_path": "/gpfswork/rech/six/commun/bigscience-training/merged-meg-ds_v2/nigercongo-st/bigscience-catalogue-data-dev_byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles_nigercongo-st_text_document",
+    "ratio": 7.211293983608406e-07
+    },
+    {
+    "dataset_path": "/gpfswork/rech/six/commun/bigscience-training/merged-meg-ds_v2/nigercongo-ny/bigscience-catalogue-data-dev_byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles_nigercongo-ny_text_document",
+    "ratio": 1.0802142612678056e-06
+    },
+    {
+    "dataset_path": "/gpfswork/rech/six/commun/bigscience-training/merged-meg-ds_v2/nigercongo-tw/bigscience-catalogue-data-dev_byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles_nigercongo-tw_text_document",
+    "ratio": 1.254210185754728e-06
+    },
+    {
+    "dataset_path": "/gpfswork/rech/six/commun/bigscience-training/merged-meg-ds_v2/nigercongo-tn/bigscience-catalogue-data-dev_byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles_nigercongo-tn_text_document",
+    "ratio": 1.4345976209947068e-06
+    },
+    {
+    "dataset_path": "/gpfswork/rech/six/commun/bigscience-training/merged-meg-ds_v2/nigercongo-ln/bigscience-catalogue-data-dev_byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles_nigercongo-ln_text_document",
+    "ratio": 1.5615957880343798e-06
+    },
+    {
+    "dataset_path": "/gpfswork/rech/six/commun/bigscience-training/merged-meg-ds_v2/nigercongo-nso/bigscience-catalogue-data-dev_byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles_nigercongo-nso_text_document",
+    "ratio": 1.5685487152725018e-06
+    },
+    {
+    "dataset_path": "/gpfswork/rech/six/commun/bigscience-training/merged-meg-ds_v2/nigercongo-fon/bigscience-catalogue-data-dev_byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles_nigercongo-fon_text_document",
+    "ratio": 2.4181733912998574e-06
+    },
+    {
+    "dataset_path": "/gpfswork/rech/six/commun/bigscience-training/merged-meg-ds_v2/nigercongo-rn/bigscience-catalogue-data-dev_byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles_nigercongo-rn_text_document",
+    "ratio": 2.6240721674330045e-06
+    },
+    {
+    "dataset_path": "/gpfswork/rech/six/commun/bigscience-training/merged-meg-ds_v2/nigercongo-wo/bigscience-catalogue-data-dev_byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles_nigercongo-wo_text_document",
+    "ratio": 3.788035372978134e-06
+    },
+    {
+    "dataset_path": "/gpfswork/rech/six/commun/bigscience-training/merged-meg-ds_v2/nigercongo-lg/bigscience-catalogue-data-dev_byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles_nigercongo-lg_text_document",
+    "ratio": 4.411728781061637e-06
+    },
+    {
+    "dataset_path": "/gpfswork/rech/six/commun/bigscience-training/merged-meg-ds_v2/nigercongo-sn/bigscience-catalogue-data-dev_byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles_nigercongo-sn_text_document",
+    "ratio": 5.462081443441739e-06
+    },
+    {
+    "dataset_path": "/gpfswork/rech/six/commun/bigscience-training/merged-meg-ds_v2/nigercongo-zu/bigscience-catalogue-data-dev_byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles_nigercongo-zu_text_document",
+    "ratio": 7.960494206222536e-06
+    },
+    {
+    "dataset_path": "/gpfswork/rech/six/commun/bigscience-training/merged-meg-ds_v2/nigercongo-ig/bigscience-catalogue-data-dev_byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles_nigercongo-ig_text_document",
+    "ratio": 1.1476452755243308e-05
+    },
+    {
+    "dataset_path": "/gpfswork/rech/six/commun/bigscience-training/merged-meg-ds_v2/nigercongo-xh/bigscience-catalogue-data-dev_byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles_nigercongo-xh_text_document",
+    "ratio": 1.420376509569156e-05
+    },
+    {
+    "dataset_path": "/gpfswork/rech/six/commun/bigscience-training/merged-meg-ds_v2/nigercongo-rw/bigscience-catalogue-data-dev_byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles_nigercongo-rw_text_document",
+    "ratio": 3.211033678381048e-05
+    },
+    {
+    "dataset_path": "/gpfswork/rech/six/commun/bigscience-training/merged-meg-ds_v2/nigercongo-yo/bigscience-catalogue-data-dev_byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles_nigercongo-yo_text_document",
+    "ratio": 5.92174680475419e-05
+    },
+    {
+    "dataset_path": "/gpfswork/rech/six/commun/bigscience-training/merged-meg-ds_v2/indic-as/bigscience-catalogue-data-dev_byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles_indic-as_text_document",
+    "ratio": 0.00011018118094953408
+    },
+    {
+    "dataset_path": "/gpfswork/rech/six/commun/bigscience-training/merged-meg-ds_v2/nigercongo-sw/bigscience-catalogue-data-dev_byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles_nigercongo-sw_text_document",
+    "ratio": 0.00016215947420710153
+    },
+    {
+    "dataset_path": "/gpfswork/rech/six/commun/bigscience-training/merged-meg-ds_v2/indic-or/bigscience-catalogue-data-dev_byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles_indic-or_text_document",
+    "ratio": 0.00035919185161466504
+    },
+    {
+    "dataset_path": "/gpfswork/rech/six/commun/bigscience-training/merged-meg-ds_v2/indic-gu/bigscience-catalogue-data-dev_byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles_indic-gu_text_document",
+    "ratio": 0.0004020242586250887
+    },
+    {
+    "dataset_path": "/gpfswork/rech/six/commun/bigscience-training/merged-meg-ds_v2/indic-mr/bigscience-catalogue-data-dev_byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles_indic-mr_text_document",
+    "ratio": 0.000501086367279419
+    },
+    {
+    "dataset_path": "/gpfswork/rech/six/commun/bigscience-training/merged-meg-ds_v2/indic-pa/bigscience-catalogue-data-dev_byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles_indic-pa_text_document",
+    "ratio": 0.0005083234921794582
+    },
+    {
+    "dataset_path": "/gpfswork/rech/six/commun/bigscience-training/merged-meg-ds_v2/zht/bigscience-catalogue-data-dev_byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles_zht_text_document",
+    "ratio": 0.0005175821828332756
+    },
+    {
+    "dataset_path": "/gpfswork/rech/six/commun/bigscience-training/merged-meg-ds_v2/indic-kn/bigscience-catalogue-data-dev_byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles_indic-kn_text_document",
+    "ratio": 0.0006189102835836872
+    },
+    {
+    "dataset_path": "/gpfswork/rech/six/commun/bigscience-training/merged-meg-ds_v2/indic-ne/bigscience-catalogue-data-dev_byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles_indic-ne_text_document",
+    "ratio": 0.0006671231017737995
+    },
+    {
+    "dataset_path": "/gpfswork/rech/six/commun/bigscience-training/merged-meg-ds_v2/indic-te/bigscience-catalogue-data-dev_byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles_indic-te_text_document",
+    "ratio": 0.0009127953712102532
+    },
+    {
+    "dataset_path": "/gpfswork/rech/six/commun/bigscience-training/merged-meg-ds_v2/indic-ml/bigscience-catalogue-data-dev_byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles_indic-ml_text_document",
+    "ratio": 0.0010332975205379394
+    },
+    {
+    "dataset_path": "/gpfswork/rech/six/commun/bigscience-training/merged-meg-ds_v2/indic-ur/bigscience-catalogue-data-dev_byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles_indic-ur_text_document",
+    "ratio": 0.0012451803025486932
+    },
+    {
+    "dataset_path": "/gpfswork/rech/six/commun/bigscience-training/merged-meg-ds_v2/eu/bigscience-catalogue-data-dev_byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles_eu_text_document",
+    "ratio": 0.0015592580929764492
+    },
+    {
+    "dataset_path": "/gpfswork/rech/six/commun/bigscience-training/merged-meg-ds_v2/indic-ta/bigscience-catalogue-data-dev_byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles_indic-ta_text_document",
+    "ratio": 0.002113263681784288
+    },
+    {
+    "dataset_path": "/gpfswork/rech/six/commun/bigscience-training/merged-meg-ds_v2/indic-bn/bigscience-catalogue-data-dev_byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles_indic-bn_text_document",
+    "ratio": 0.005491432072714383
+    },
+    {
+    "dataset_path": "/gpfswork/rech/six/commun/bigscience-training/merged-meg-ds_v2/indic-hi/bigscience-catalogue-data-dev_byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles_indic-hi_text_document",
+    "ratio": 0.007468936340601538
+    },
+    {
+    "dataset_path": "/gpfswork/rech/six/commun/bigscience-training/merged-meg-ds_v2/id/bigscience-catalogue-data-dev_byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles_id_text_document",
+    "ratio": 0.010918042445283764
+    },
+    {
+    "dataset_path": "/gpfswork/rech/six/commun/bigscience-training/merged-meg-ds_v2/ca/bigscience-catalogue-data-dev_byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles_ca_text_document",
+    "ratio": 0.011240177446143296
+    },
+    {
+    "dataset_path": "/gpfswork/rech/six/commun/bigscience-training/merged-meg-ds_v2/vi/bigscience-catalogue-data-dev_byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles_vi_text_document",
+    "ratio": 0.024618602569261978
+    },
+    {
+    "dataset_path": "/gpfswork/rech/six/commun/bigscience-training/merged-meg-ds_v2/ar/bigscience-catalogue-data-dev_byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles_ar_text_document",
+    "ratio": 0.03306323144550797
+    },
+    {
+    "dataset_path": "/gpfswork/rech/six/commun/bigscience-training/merged-meg-ds_v2/pt/bigscience-catalogue-data-dev_byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles_pt_text_document",
+    "ratio": 0.049537176530748885
+    },
+    {
+    "dataset_path": "/gpfswork/rech/six/commun/bigscience-training/merged-meg-ds_v2/oscar_zh/bigscience-catalogue-data-dev_byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles_oscar_zh_text_document",
+    "ratio": 0.0553990566108835
+    },
+    {
+    "dataset_path": "/gpfswork/rech/six/commun/bigscience-training/merged-meg-ds_v2/oscar_en/bigscience-catalogue-data-dev_byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles_oscar_en_text_document",
+    "ratio": 0.08138409077754236
+    },
+    {
+    "dataset_path": "/gpfswork/rech/six/commun/bigscience-training/merged-meg-ds_v2/es/bigscience-catalogue-data-dev_byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles_es_text_document",
+    "ratio": 0.10667605099879819
+    },
+    {
+    "dataset_path": "/gpfswork/rech/six/commun/bigscience-training/merged-meg-ds_v2/zhs/bigscience-catalogue-data-dev_byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles_zhs_text_document",
+    "ratio": 0.12089121520420677
+    },
+    {
+    "dataset_path": "/gpfswork/rech/six/commun/bigscience-training/merged-meg-ds_v2/code/bigscience-catalogue-data-dev_byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles_code_text_document",
+    "ratio": 0.13024371838857235
+    },
+    {
+    "dataset_path": "/gpfswork/rech/six/commun/bigscience-training/merged-meg-ds_v2/fr/bigscience-catalogue-data-dev_byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles_fr_text_document",
+    "ratio": 0.13054073905636543
+    },
+    {
+    "dataset_path": "/gpfswork/rech/six/commun/bigscience-training/merged-meg-ds_v2/en/bigscience-catalogue-data-dev_byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles_en_text_document",
+    "ratio": 0.2216634990836396
+    }
+]

data/catalogue/training_dataset_ratios_merged_nigercongo.json

@@ -0,0 +1,114 @@
+[
+  {
+    "dataset_path": "/gpfswork/rech/six/commun/bigscience-training/merged-meg-ds_v2/indic-as/bigscience-catalogue-data-dev_byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles_indic-as_text_document",
+    "ratio": 0.00011018118094953408
+  },
+  {
+    "dataset_path": "/gpfswork/rech/six/commun/bigscience-training/merged-meg-ds_v2/indic-or/bigscience-catalogue-data-dev_byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles_indic-or_text_document",
+    "ratio": 0.00035919185161466504
+  },
+  {
+    "dataset_path": "/gpfswork/rech/six/commun/bigscience-training/merged-meg-ds_v2/indic-gu/bigscience-catalogue-data-dev_byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles_indic-gu_text_document",
+    "ratio": 0.0004020242586250887
+  },
+  {
+    "dataset_path": "/gpfswork/rech/six/commun/bigscience-training/merged-meg-ds_v2/indic-mr/bigscience-catalogue-data-dev_byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles_indic-mr_text_document",
+    "ratio": 0.000501086367279419
+  },
+  {
+    "dataset_path": "/gpfswork/rech/six/commun/bigscience-training/merged-meg-ds_v2/indic-pa/bigscience-catalogue-data-dev_byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles_indic-pa_text_document",
+    "ratio": 0.0005083234921794582
+  },
+  {
+    "dataset_path": "/gpfswork/rech/six/commun/bigscience-training/merged-meg-ds_v2/zht/bigscience-catalogue-data-dev_byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles_zht_text_document",
+    "ratio": 0.0005175821828332756
+  },
+  {
+    "dataset_path": "/gpfswork/rech/six/commun/bigscience-training/merged-meg-ds_v2/indic-kn/bigscience-catalogue-data-dev_byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles_indic-kn_text_document",
+    "ratio": 0.0006189102835836872
+  },
+  {
+    "dataset_path": "/gpfswork/rech/six/commun/bigscience-training/merged-meg-ds_v2/indic-ne/bigscience-catalogue-data-dev_byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles_indic-ne_text_document",
+    "ratio": 0.0006671231017737995
+  },
+  {
+    "dataset_path": "/gpfswork/rech/six/commun/bigscience-training/merged-meg-ds_v2/indic-te/bigscience-catalogue-data-dev_byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles_indic-te_text_document",
+    "ratio": 0.0009127953712102532
+  },
+  {
+    "dataset_path": "/gpfswork/rech/six/commun/bigscience-training/merged-meg-ds_v2/indic-ml/bigscience-catalogue-data-dev_byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles_indic-ml_text_document",
+    "ratio": 0.0010332975205379394
+  },
+  {
+    "dataset_path": "/gpfswork/rech/six/commun/bigscience-training/merged-meg-ds_v2/indic-ur/bigscience-catalogue-data-dev_byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles_indic-ur_text_document",
+    "ratio": 0.0012451803025486932
+  },
+  {
+    "dataset_path": "/gpfswork/rech/six/commun/bigscience-training/merged-meg-ds_v2/eu/bigscience-catalogue-data-dev_byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles_eu_text_document",
+    "ratio": 0.0015592580929764492
+  },
+  {
+    "dataset_path": "/gpfswork/rech/six/commun/bigscience-training/merged-meg-ds_v2/indic-ta/bigscience-catalogue-data-dev_byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles_indic-ta_text_document",
+    "ratio": 0.002113263681784288
+  },
+  {
+    "dataset_path": "/gpfswork/rech/six/commun/bigscience-training/merged-meg-ds_v2/indic-bn/bigscience-catalogue-data-dev_byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles_indic-bn_text_document",
+    "ratio": 0.005491432072714383
+  },
+  {
+    "dataset_path": "/gpfswork/rech/six/commun/bigscience-training/merged-meg-ds_v2/indic-hi/bigscience-catalogue-data-dev_byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles_indic-hi_text_document",
+    "ratio": 0.007468936340601538
+  },
+  {
+    "dataset_path": "/gpfswork/rech/six/commun/bigscience-training/merged-meg-ds_v2/id/bigscience-catalogue-data-dev_byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles_id_text_document",
+    "ratio": 0.010918042445283764
+  },
+  {
+    "dataset_path": "/gpfswork/rech/six/commun/bigscience-training/merged-meg-ds_v2/ca/bigscience-catalogue-data-dev_byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles_ca_text_document",
+    "ratio": 0.011240177446143296
+  },
+  {
+    "dataset_path": "/gpfswork/rech/six/commun/bigscience-training/merged-meg-ds_v2/vi/bigscience-catalogue-data-dev_byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles_vi_text_document",
+    "ratio": 0.024618602569261978
+  },
+  {
+    "dataset_path": "/gpfswork/rech/six/commun/bigscience-training/merged-meg-ds_v2/ar/bigscience-catalogue-data-dev_byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles_ar_text_document",
+    "ratio": 0.03306323144550797
+  },
+  {
+    "dataset_path": "/gpfswork/rech/six/commun/bigscience-training/merged-meg-ds_v2/pt/bigscience-catalogue-data-dev_byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles_pt_text_document",
+    "ratio": 0.049537176530748885
+  },
+  {
+    "dataset_path": "/gpfswork/rech/six/commun/bigscience-training/merged-meg-ds_v2/oscar_zh/bigscience-catalogue-data-dev_byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles_oscar_zh_text_document",
+    "ratio": 0.0553990566108835
+  },
+  {
+    "dataset_path": "/gpfswork/rech/six/commun/bigscience-training/merged-meg-ds_v2/oscar_en/bigscience-catalogue-data-dev_byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles_oscar_en_text_document",
+    "ratio": 0.08138409077754236
+  },
+  {
+    "dataset_path": "/gpfswork/rech/six/commun/bigscience-training/merged-meg-ds_v2/es/bigscience-catalogue-data-dev_byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles_es_text_document",
+    "ratio": 0.10667605099879819
+  },
+  {
+    "dataset_path": "/gpfswork/rech/six/commun/bigscience-training/merged-meg-ds_v2/zhs/bigscience-catalogue-data-dev_byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles_zhs_text_document",
+    "ratio": 0.12089121520420677
+  },
+  {
+    "dataset_path": "/gpfswork/rech/six/commun/bigscience-training/merged-meg-ds_v2/code/bigscience-catalogue-data-dev_byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles_code_text_document",
+    "ratio": 0.13024371838857235
+  },
+  {
+    "dataset_path": "/gpfswork/rech/six/commun/bigscience-training/merged-meg-ds_v2/fr/bigscience-catalogue-data-dev_byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles_fr_text_document",
+    "ratio": 0.13054073905636543
+  },
+  {
+    "dataset_path": "/gpfswork/rech/six/commun/bigscience-training/merged-meg-ds_v2/en/bigscience-catalogue-data-dev_byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles_en_text_document",
+    "ratio": 0.2216634990836396
+  },
+  {
+    "dataset_path": "/gpfswork/rech/six/commun/bigscience-training/merged-meg-ds_v2/nigercongo/bigscience-catalogue-data-dev_byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles_nigercongo_text_document",
+    "ratio": 0.00031581334183345284
+  }
+]

data/catalogue/training_dataset_ratios_merged_nigercongo_v3.json

@@ -0,0 +1,114 @@
+[
+  {
+    "dataset_path": "/gpfswork/rech/six/commun/bigscience-training/merged-meg-ds_v3_pii/ar/bigscience-catalogue-data-dev_byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles_ar_text_document",
+    "ratio": 0.0330676168743166
+  },
+  {
+    "dataset_path": "/gpfswork/rech/six/commun/bigscience-training/merged-meg-ds_v3_pii/ca/bigscience-catalogue-data-dev_byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles_ca_text_document",
+    "ratio": 0.011242051312222764
+  },
+  {
+    "dataset_path": "/gpfswork/rech/six/commun/bigscience-training/merged-meg-ds_v3_pii/code/bigscience-catalogue-data-dev_byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles_code_text_document",
+    "ratio": 0.13027200903379185
+  },
+  {
+    "dataset_path": "/gpfswork/rech/six/commun/bigscience-training/merged-meg-ds_v3_pii/en/bigscience-catalogue-data-dev_byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles_en_text_document",
+    "ratio": 0.22171164529099704
+  },
+  {
+    "dataset_path": "/gpfswork/rech/six/commun/bigscience-training/merged-meg-ds_v3_pii/es/bigscience-catalogue-data-dev_byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles_es_text_document",
+    "ratio": 0.10667815627928671
+  },
+  {
+    "dataset_path": "/gpfswork/rech/six/commun/bigscience-training/merged-meg-ds_v3_pii/eu/bigscience-catalogue-data-dev_byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles_eu_text_document",
+    "ratio": 0.0015595123898173287
+  },
+  {
+    "dataset_path": "/gpfswork/rech/six/commun/bigscience-training/merged-meg-ds_v3_pii/fr/bigscience-catalogue-data-dev_byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles_fr_text_document",
+    "ratio": 0.13054018439603915
+  },
+  {
+    "dataset_path": "/gpfswork/rech/six/commun/bigscience-training/merged-meg-ds_v3_pii/id/bigscience-catalogue-data-dev_byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles_id_text_document",
+    "ratio": 0.01091803753667153
+  },
+  {
+    "dataset_path": "/gpfswork/rech/six/commun/bigscience-training/merged-meg-ds_v3_pii/indic-as/bigscience-catalogue-data-dev_byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles_indic-as_text_document",
+    "ratio": 0.00011021422347108609
+  },
+  {
+    "dataset_path": "/gpfswork/rech/six/commun/bigscience-training/merged-meg-ds_v3_pii/indic-bn/bigscience-catalogue-data-dev_byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles_indic-bn_text_document",
+    "ratio": 0.005492381453597748
+  },
+  {
+    "dataset_path": "/gpfswork/rech/six/commun/bigscience-training/merged-meg-ds_v3_pii/indic-gu/bigscience-catalogue-data-dev_byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles_indic-gu_text_document",
+    "ratio": 0.0004021215011318779
+  },
+  {
+    "dataset_path": "/gpfswork/rech/six/commun/bigscience-training/merged-meg-ds_v3_pii/indic-hi/bigscience-catalogue-data-dev_byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles_indic-hi_text_document",
+    "ratio": 0.007470068593492175
+  },
+  {
+    "dataset_path": "/gpfswork/rech/six/commun/bigscience-training/merged-meg-ds_v3_pii/indic-kn/bigscience-catalogue-data-dev_byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles_indic-kn_text_document",
+    "ratio": 0.0006190467776576425
+  },
+  {
+    "dataset_path": "/gpfswork/rech/six/commun/bigscience-training/merged-meg-ds_v3_pii/indic-ml/bigscience-catalogue-data-dev_byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles_indic-ml_text_document",
+    "ratio": 0.0010335296343329384
+  },
+  {
+    "dataset_path": "/gpfswork/rech/six/commun/bigscience-training/merged-meg-ds_v3_pii/indic-mr/bigscience-catalogue-data-dev_byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles_indic-mr_text_document",
+    "ratio": 0.0005012010684646179
+  },
+  {
+    "dataset_path": "/gpfswork/rech/six/commun/bigscience-training/merged-meg-ds_v3_pii/indic-ne/bigscience-catalogue-data-dev_byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles_indic-ne_text_document",
+    "ratio": 0.0006672772956128299
+  },
+  {
+    "dataset_path": "/gpfswork/rech/six/commun/bigscience-training/merged-meg-ds_v3_pii/indic-or/bigscience-catalogue-data-dev_byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles_indic-or_text_document",
+    "ratio": 0.00035928138344705506
+  },
+  {
+    "dataset_path": "/gpfswork/rech/six/commun/bigscience-training/merged-meg-ds_v3_pii/indic-pa/bigscience-catalogue-data-dev_byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles_indic-pa_text_document",
+    "ratio": 0.0005084433130291778
+  },
+  {
+    "dataset_path": "/gpfswork/rech/six/commun/bigscience-training/merged-meg-ds_v3_pii/indic-ta/bigscience-catalogue-data-dev_byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles_indic-ta_text_document",
+    "ratio": 0.0021137328219915496
+  },
+  {
+    "dataset_path": "/gpfswork/rech/six/commun/bigscience-training/merged-meg-ds_v3_pii/indic-te/bigscience-catalogue-data-dev_byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles_indic-te_text_document",
+    "ratio": 0.0009129946225980253
+  },
+  {
+    "dataset_path": "/gpfswork/rech/six/commun/bigscience-training/merged-meg-ds_v3_pii/indic-ur/bigscience-catalogue-data-dev_byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles_indic-ur_text_document",
+    "ratio": 0.0012454301613725426
+  },
+  {
+    "dataset_path": "/gpfswork/rech/six/commun/bigscience-training/merged-meg-ds_v3_pii/nigercongo-all/bigscience-catalogue-data-dev_byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles_nigercongo-all_text_document",
+    "ratio": 0.00031588689199263235
+  },
+  {
+    "dataset_path": "/gpfswork/rech/six/commun/bigscience-training/merged-meg-ds_v3_pii/oscar-en/meg_ds_bigscience-catalogue-data-dev_byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles_text_document",
+    "ratio": 0.08137213783015229
+  },
+  {
+    "dataset_path": "/gpfswork/rech/six/commun/bigscience-training/merged-meg-ds_v3_pii/oscar-zh/meg_ds_bigscience-catalogue-data-dev_byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles_text_document",
+    "ratio": 0.055293935695898196
+  },
+  {
+    "dataset_path": "/gpfswork/rech/six/commun/bigscience-training/merged-meg-ds_v3_pii/pt/bigscience-catalogue-data-dev_byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles_pt_text_document",
+    "ratio": 0.04954150576361177
+  },
+  {
+    "dataset_path": "/gpfswork/rech/six/commun/bigscience-training/merged-meg-ds_v3_pii/vi/bigscience-catalogue-data-dev_byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles_vi_text_document",
+    "ratio": 0.02461641286531197
+  },
+  {
+    "dataset_path": "/gpfswork/rech/six/commun/bigscience-training/merged-meg-ds_v3_pii/zhs/bigscience-catalogue-data-dev_byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles_zhs_text_document",
+    "ratio": 0.12091748245519074
+  },
+  {
+    "dataset_path": "/gpfswork/rech/six/commun/bigscience-training/merged-meg-ds_v3_pii/zht/bigscience-catalogue-data-dev_byte-level-bpe-tokenizer-no-norm-250k-whitespace-and-eos-regex-alpha-v3-dedup-lines-articles_zht_text_document",
+    "ratio": 0.0005177025345001541
+  }
+]

data/mc4/README.md

@@ -0,0 +1,26 @@
+# mc4
+
+## Megatron pre-processed files
+
+
+These are the megatron-ready mc4 files:
+
+- 1.3TB: `$six_ALL_CCFRWORK/datasets-custom/mc4/mc4_preprocessing`
+
+Should something get corrupted there is a backup:
+
+- 1.3TB: `$six_ALL_CCFRSTORE/datasets-custom/mc4/mc4_preprocessing`
+
+If files need to re-pre-processed, the original jsonl files are at:
+
+- 186GB: `$six_ALL_CCFRSTORE/datasets-custom/mc4/mc4_sampled_raw`
+
+
+## How pre-processing was done
+
+The pre-processing was done outside of JZ, and was downloaded from:
+
+* [mc4_preprocessing](https://console.cloud.google.com/storage/browser/bigscience/mc4_preprocessing?pageState=(%22StorageObjectListTable%22:(%22f%22:%22%255B%255D%22)))
+* [mc4_sampled_raw](https://console.cloud.google.com/storage/browser/bigscience/mc4_sampled_raw?pageState=(%22StorageObjectListTable%22:(%22f%22:%22%255B%255D%22)))
+
+To download one needs to activate the already installed on JZ [google-cloud-sdk](../../jz/tools/google-cloud-sdk.md) and then use `gsutil` as instructed at the `Download` tab in the links above.

data/openwebtext/openwebtext-to-jsonl.py

@@ -0,0 +1,28 @@
+#!/usr/bin/env python
+
+# generate jsonl version of dataset that can be fed to megatron-lm preprocessor
+#
+# full dataset
+# ./openwebtext-to-jsonl.py
+#
+# 10k small dataset
+# ./openwebtext-to-jsonl.py -10k
+
+import sys
+from datasets import load_dataset
+
+if "-10k" in sys.argv:
+    dataset_name = "stas/openwebtext-10k"
+else:
+    dataset_name = "openwebtext"
+
+name = dataset_name.split('/')[-1]
+ds = load_dataset(dataset_name, split='train')
+ds.to_json(f"{name}.jsonl", orient="records", lines=True)
+
+# subset to jsonlines
+# n_samples = 1000
+# ds = load_dataset(dataset_name, split='train')
+# ds_small = ds.select(range(n_samples))
+# path = f"{dataset_name}-{n_samples}.jsonl"
+# ds_small.to_json(path, orient="records", lines=True)

data/oscar-multilingual/README.md

@@ -0,0 +1,218 @@
+# OSCAR
+
+
+## Megatron pre-processed files
+
+These are the megatron-ready OSCAR files:
+
+- Full 300M version (529GB) : `$six_ALL_CCFRWORK/datasets-custom/oscar-en`
+- Tiny 10K version (56M): `$six_ALL_CCFRWORK/datasets-custom/oscar-en-10k`
+
+Each folder contains: `meg-gpt2_text_document.bin` and `meg-gpt2_text_document.idx` and Megatron-LM training script expects the following argument:
+```
+--data-path $six_ALL_CCFRWORK/datasets-custom/oscar-en/meg-gpt2_text_document
+```
+
+Should something get corrupted there is a backup:
+
+- Full 300M version (529GB) : `$six_ALL_CCFRSTORE/datasets-custom/oscar-en`
+- Tiny 10K version (56M): `$six_ALL_CCFRSTORE/datasets-custom/oscar-en-10k`
+
+
+
+
+## How pre-processing was done
+
+In general the process is to first generate jsonl version of the dataset, while filtering out entries smaller than 1K, and then run that jsonl data through Megatron-LM preprocessing tool.
+
+The rest of this document is the step by step process of accomplishing that in an efficient way.
+
+**Update: Now that we better understand Megatron-LM's dataloader we know that it contacts all docs on the fly and delivers seqlen at a time as a single sample ([reference](https://github.com/NVIDIA/Megatron-LM/blob/90e0a0dd08159e1c95f4f9d99bb8687f327d36c3/megatron/data/gpt_dataset.py#L169-L185). So we don't need to filter out docs that are shorter than seqlen. Therefore in the future runs. We should adjust `oscar-to-jsonl.py` to remove the filtering.**
+
+1. Convert `datasets` to `jsonl` which is the format required by Megatron-LM
+
+The main script is [oscar-to-jsonl.py](./oscar-multilingual-to-jsonl.py). Edit to change languages to use, initially using just English.
+
+Note, that since shuffling slows the writeout process by 5-7 times, we don't shuffle in the script, but post-process it externally. See step 3.
+
+To launch: [oscar-to-jsonl.slurm](./oscar-multilingual-to-jsonl.slurm).
+
+With "unshuffled_deduplicated_en" after filtering large entries (`>=1024`) we end up with 70754K examples out of 304230K total (about 1/4th of the full dataset).
+
+The result is 5 files `oscar-[0-4].jsonl` of about 250GB each.
+
+Runtime: 2-3h to download, ~2h to build, ~8h to filter, ~1.5h to write shards out
+
+
+2. Concatenate
+
+```
+cat oscar-[0-4].jsonl > oscar-en.jsonl
+```
+
+This gives us a 1.2TB file.
+
+Check:
+```
+$ wc -l oscar-en.jsonl
+304230423 oscar-en.jsonl
+```
+
+Runtime: a few minutes
+
+
+
+3. Shuffle
+
+Megatron requires users to do their own shuffling of jsonl input.
+
+It was too slow to do inside the filtering script, so we are using a post-processing solution.
+Using https://github.com/alexandres/terashuf and 150GB RAM in ~1.5h we shuffle the file.
+
+Important: note that the slurm job uses SCRATCH for `TMPDIR` and also sets the memory limit it can use to 150.0 (GB) (slightly under 160GB available on this slurm allocation to allow for other processes).
+
+To launch: [oscar-fast-shuffle.slurm](./oscar-fast-shuffle.slurm)
+
+`terashuf` is in `$six_ALL_CCFRWORK/bin/terashuf`
+
+The result is `oscar-shuffled.jsonl`
+
+Runtime: 2h
+
+
+
+4. Megatron-LM preprocess
+
+**Update**: that was an error, we can actually run for 100h on `-p cpu_p1` and so the normal script can complete no problem, but as a result of this mistake we can now pre-process data much faster.
+
+We only have 20h to do processing which is not enough to process 300M records. Trying to do the whole thing in one preprocessing script took more than 24h and thus failed. Adding more than 16 workers didn't speed things up.
+
+So we are splitting it in 4 chunks of ~80M records
+
+```
+split -l 77000000 oscar-en-shuffled.jsonl oscar
+mv oscaraa oscar-en-shuffled-p1.jsonl
+mv oscarab oscar-en-shuffled-p2.jsonl
+mv oscarac oscar-en-shuffled-p3.jsonl
+mv oscarad oscar-en-shuffled-p4.jsonl
+```
+
+We do the pre-processing:
+
+The main script to launch: [oscar-jsonl-to-meg-gpt2.slurm](./oscar-jsonl-to-meg.slurm), and we need to make copies of it for each chunk:
+
+```
+cp oscar-jsonl-to-meg-gpt2.slurm oscar-jsonl-to-meg-gpt2-1.slurm
+cp oscar-jsonl-to-meg-gpt2.slurm oscar-jsonl-to-meg-gpt2-2.slurm
+cp oscar-jsonl-to-meg-gpt2.slurm oscar-jsonl-to-meg-gpt2-3.slurm
+cp oscar-jsonl-to-meg-gpt2.slurm oscar-jsonl-to-meg-gpt2-4.slurm
+perl -pi -e 's|p1|p1|' oscar-jsonl-to-meg-gpt2-1.slurm
+perl -pi -e 's|p1|p2|' oscar-jsonl-to-meg-gpt2-2.slurm
+perl -pi -e 's|p1|p3|' oscar-jsonl-to-meg-gpt2-3.slurm
+perl -pi -e 's|p1|p4|' oscar-jsonl-to-meg-gpt2-4.slurm
+```
+
+```
+sbatch oscar-jsonl-to-meg-gpt2-1.slurm
+sbatch oscar-jsonl-to-meg-gpt2-2.slurm
+sbatch oscar-jsonl-to-meg-gpt2-3.slurm
+sbatch oscar-jsonl-to-meg-gpt2-4.slurm
+```
+
+This took about 6h each but run in parallel on different instances. This is surprisingly the projected time for the initial attempt to run in in one chunk, which was projected to 24 hours, and couldn't fit into 20h cap. So we finished the whole thing in 6 hours.
+
+Outcome:
+
+```
+$ ls -1sh meg-gpt2-p*
+131G meg-gpt2-p1_text_document.bin
+1.4G meg-gpt2-p1_text_document.idx
+131G meg-gpt2-p2_text_document.bin
+1.4G meg-gpt2-p2_text_document.idx
+131G meg-gpt2-p3_text_document.bin
+1.4G meg-gpt2-p3_text_document.idx
+138G meg-gpt2-p4_text_document.bin
+1.5G meg-gpt2-p4_text_document.idx
+```
+
+Next merging: [oscar-meg-gpt2-merge.slurm](./oscar-meg-gpt2-merge.slurm)
+
+Runtime: 22min - needed 26GB RSS RAM
+
+Outcome: 304_230_423 records
+
+```
+$ ls -1sh meg-gpt2_text_document.*
+529G meg-gpt2_text_document.bin
+5.7G meg-gpt2_text_document.idx
+```
+
+Total runtime: under 7h.
+
+Let's also make a small 10k version for experiments:
+
+```
+head -10000 oscar-shuffled.jsonl > oscar-shuffled-10k.jsonl
+```
+and then process with the same slurm script above, but changing the input to `oscar-shuffled-10k.jsonl`
+
+
+
+5. Final destination
+
+We did all the processing on the SCRATCH partition which gets wiped out every 30 days, so we need to move the files to where they will not be deleted.
+
+Since at this moment we used just the English part of the OSCAR dataset, let's include that in the folder name to differentiate from other builds that will be multi-lingual.
+
+Make the final result which will be used by the megatron training script available on the persistent WORK partition:
+
+```
+mkdir oscar-en
+mv meg-gpt2_text_document.* oscar-en
+cp -r oscar-en $six_ALL_CCFRWORK/datasets-custom
+```
+
+Back it up to STORE:
+
+It's already binary and just 2 files, so no need to tar (STORE has limited inodes)
+```
+mkdir -p $six_ALL_CCFRSTORE/datasets-custom
+cp -r oscar-en $six_ALL_CCFRSTORE/datasets-custom
+```
+
+Also copy the small version for experiments to WORK and STORE:
+```
+cp -r oscar-en-10k $six_ALL_CCFRWORK/datasets-custom
+cp -r oscar-en-10k $six_ALL_CCFRSTORE/datasets-custom
+```
+
+Tar/gz `oscar-shuffled.jsonl` and the dataset files to STORE:
+
+```
+
+
+```
+
+6. Estimate total number of tokens
+
+Make a 1GB slice:
+```
+$ head -79000 oscar-en-shuffled.jsonl > oscar-1GB.jsonl
+$ ls -sh oscar-1GB.jsonl
+1.0G oscar-1GB.jsonl
+```
+
+Analyze it (low mem-footprint):
+```
+$ python -c "import json, sys; \
+from transformers import GPT2TokenizerFast; \
+tokenizer = GPT2TokenizerFast.from_pretrained('gpt2'); \
+print(sum(tokenizer(json.loads(l)['text'], return_length=True).length[0] for l in sys.stdin.readlines()))" < oscar-1GB.jsonl
+234260484
+```
+
+Extrapolate:
+
+Thus 234M tokens in 1GB, ~280B tokens in 1.2TB (`234*1200`)
+
+Incidentally this coincides with @Yozh's `FILE_SIZE_IN_GBS/4.5` formula! (average 4.5chars per word)

data/oscar-multilingual/download-oscars.py

@@ -0,0 +1,50 @@
+#!/usr/bin/env python
+#
+# generate jsonl version of dataset that can be fed to megatron-lm pre-processor
+#
+# see various notes in the scripts for different options
+#
+# full dataset:
+# ./oscar-multilingual-to-jsonl.py
+# cat oscar-[0-4].jsonl > oscar.jsonl
+#
+# small dataset (0.1%):
+# ./oscar-multilingual-to-jsonl.py -s
+# cat oscar-[0-4].jsonl > oscar.jsonl
+
+import logging
+import os
+
+import datasets
+
+print(f"Using datasets=={datasets.__version__}")
+
+DATASET_NAME = "oscar"
+logging.getLogger("transformers.tokenization_utils_base").setLevel(logging.ERROR)
+
+### Build/Load Datasets
+
+# Once this part of the process completes it gets cached, so on subsequent runs it'll be much faster
+
+language_subsets = {
+    "unshuffled_deduplicated_ar",
+    "unshuffled_deduplicated_sw",
+    "unshuffled_deduplicated_zh",
+    # "unshuffled_deduplicated_en",
+    "unshuffled_deduplicated_fr",
+    "unshuffled_deduplicated_pt",
+    "unshuffled_deduplicated_es",
+    "unshuffled_deduplicated_ja",
+    "unshuffled_deduplicated_ru",
+    "unshuffled_deduplicated_hi",
+    "unshuffled_deduplicated_ur",
+    "unshuffled_deduplicated_bn",
+    "unshuffled_deduplicated_id",
+    "unshuffled_deduplicated_am",
+    "unshuffled_deduplicated_ca",
+}
+
+for language_subset in language_subsets:
+    builder = datasets.load_dataset_builder(DATASET_NAME, language_subset, cache_dir='cache')
+    if not os.path.isdir(builder.cache_dir):
+        builder.download_and_prepare(ignore_verifications=True)

data/oscar-multilingual/download-oscars.slurm

@@ -0,0 +1,18 @@
+#!/bin/bash
+#SBATCH --job-name=download-oscars    # job name
+#SBATCH --ntasks=1                   # number of MP tasks
+#SBATCH --nodes=1
+#SBATCH --cpus-per-task=40           # number of cores per tasks
+#SBATCH --hint=nomultithread         # we get physical cores not logical
+#SBATCH --time=20:00:00             # maximum execution time (HH:MM:SS)
+#SBATCH --output=%x-%j.out           # output file name
+#SBATCH --account=six@cpu
+#SBATCH --partition=compil
+
+set -x -e
+
+source $six_ALL_CCFRWORK/start-prod
+
+# use SCRATCH for building as it's much faster
+cd $six_ALL_CCFRSCRATCH/datasets/oscar-multilingual
+python $SCRATCH/bigscience/data/oscar-multilingual/download-oscars.py

data/oscar-multilingual/oscar-fast-shuffle.slurm

@@ -0,0 +1,27 @@
+#!/bin/bash
+#SBATCH --job-name=oscar-fast-shuffle      # job name
+#SBATCH --ntasks=1                   # number of MP tasks
+#SBATCH --nodes=1                    # number of nodes
+#SBATCH --cpus-per-task=40           # number of cores per task
+#SBATCH --hint=nomultithread         # we get physical cores not logical
+#SBATCH --time=100:00:00             # maximum execution time (HH:MM:SS)
+#SBATCH --output=%x-%j.out           # output file name
+#SBATCH --account=six@cpu            # allocation account
+#SBATCH --partition=cpu_p1
+
+set -x -e
+
+source $six_ALL_CCFRWORK/start-prod
+
+# must set tmp to SCRATCH to be fast
+export TMPDIR=$six_ALL_CCFRSCRATCH/tmp
+mkdir -p $TMPDIR
+
+# memory to use in GBs float
+export MEMORY=150.0
+
+input=oscar-en.jsonl
+output=oscar-en-shuffled.jsonl
+
+cd $six_ALL_CCFRSCRATCH/datasets/oscar-small
+/usr/bin/time -v $six_ALL_CCFRWORK/bin/terashuf < $input > $output

data/oscar-multilingual/oscar-jsonl-to-meg.sh

@@ -0,0 +1,7 @@
+#!/bin/bash
+# the $1 argument can be "equal" or "alpha" to choose the corresponding tokenizer
+# no en in this script; we've mostly processed it on GCP
+for language in fr es zh hi ur bn id ca ar pt vi eu
+do
+    sbatch $ALL_CCFRWORK/code/bigscience/data/oscar-multilingual/oscar-jsonl-to-meg.slurm $language $1
+done

data/oscar-multilingual/oscar-jsonl-to-meg.slurm

@@ -0,0 +1,30 @@
+#!/bin/bash
+#SBATCH --job-name=oscar-jsonl-to-meg-equal # job name
+#SBATCH --ntasks=1                   # number of MP tasks
+#SBATCH --nodes=1
+#SBATCH --cpus-per-task=40           # number of cores per tasks
+#SBATCH --hint=nomultithread         # we get physical cores not logical
+#SBATCH --time=20:00:00             # maximum execution time (HH:MM:SS)
+#SBATCH --qos=qos_cpu-t3
+#SBATCH --output=%x-%j.out          # output file name
+#SBATCH --account=six@cpu
+#SBATCH --partition=cpu_p1
+
+set -x -e
+
+source $six_ALL_CCFRWORK/start-prod
+export HF_DATASETS_OFFLINE=1
+export TRANSFORMERS_OFFLINE=1
+
+input=$six_ALL_CCFRWORK/datasets-custom/oscar-multilingual/oscar_${1}.jsonl
+output=$six_ALL_CCFRWORK/datasets-custom/oscar-multilingual-${2}-tok/oscar_${1}_${2}
+
+cd $ALL_CCFRWORK/code/Megatron-DeepSpeed
+/usr/bin/time -v python tools/preprocess_data.py \
+       --input $input \
+       --output-prefix $output \
+       --dataset-impl mmap \
+       --tokenizer-type PretrainedFromHF \
+       --tokenizer-name-or-path bigscience/oscar_13_languages_${2}_weight \
+       --append-eod \
+       --workers 25

data/oscar-multilingual/oscar-meg-gpt2-merge.slurm

@@ -0,0 +1,24 @@
+#!/bin/bash
+#SBATCH --job-name=oscar-meg-gpt2-merge.slurm # job name
+#SBATCH --ntasks=1                   # number of MP tasks
+#SBATCH --nodes=1
+#SBATCH --cpus-per-task=10           # number of cores per task
+#SBATCH --hint=nomultithread         # we get physical cores not logical
+#SBATCH --time=100:00:00             # maximum execution time (HH:MM:SS)
+#SBATCH --output=%x-%j.out           # output file name
+#SBATCH --account=six@cpu
+#SBATCH --partition=cpu_p1
+
+set -x -e
+
+source $six_ALL_CCFRWORK/start-prod
+
+cd $six_ALL_CCFRWORK/code/Megatron-DeepSpeed
+DATA=$six_ALL_CCFRSCRATCH/datasets/oscar-multilingual
+/usr/bin/time -v python tools/merge_preprocessed_data.py \
+    --datasets \
+    $DATA/meg-gpt2-p1_text_document \
+    $DATA/meg-gpt2-p2_text_document \
+    $DATA/meg-gpt2-p3_text_document \
+    $DATA/meg-gpt2-p4_text_document \
+    --output-prefix $DATA/meg-gpt2_text_document

data/oscar-multilingual/oscar-multilingual-to-jsonl.py

@@ -0,0 +1,106 @@
+#!/usr/bin/env python
+#
+# generate jsonl version of dataset that can be fed to megatron-lm pre-processor
+#
+# see various notes in the scripts for different options
+#
+# full dataset:
+# ./oscar-multilingual-to-jsonl.py
+# cat oscar-[0-4].jsonl > oscar.jsonl
+#
+# small dataset (0.1%):
+# ./oscar-multilingual-to-jsonl.py -s
+# cat oscar-[0-4].jsonl > oscar.jsonl
+
+import logging
+from argparse import ArgumentParser
+from multiprocessing import Process, Queue
+
+from datasets import load_dataset, ReadInstruction
+
+import datasets
+
+print(f"Using datasets=={datasets.__version__}")
+
+DATASET_NAME = "oscar"
+
+logging.getLogger("transformers.tokenization_utils_base").setLevel(logging.ERROR)
+
+parser = ArgumentParser()
+parser.add_argument('-s', '--subset', action='store_true', help='Process and save a subset (0.1%) of data')
+args = parser.parse_args()
+
+# Once this part of the process runs it gets cached, so on subsequent runs it'll be much faster
+
+split = ReadInstruction("train", to=0.1 if args.subset else 100, unit="%")
+
+### Build/Load Datasets
+
+# Once this part of the process completes it gets cached, so on subsequent runs it'll be much faster
+
+language_subsets = {
+    "unshuffled_deduplicated_hi",
+    "unshuffled_deduplicated_ur",
+    "unshuffled_deduplicated_bn",
+    "unshuffled_deduplicated_id",
+    "unshuffled_deduplicated_ca",
+    "unshuffled_deduplicated_eu",
+    "unshuffled_deduplicated_ar",
+    "unshuffled_deduplicated_sw",
+    "unshuffled_deduplicated_zh",
+    "unshuffled_deduplicated_en",
+    "unshuffled_deduplicated_fr",
+    "unshuffled_deduplicated_pt",
+    "unshuffled_deduplicated_es",
+    "unshuffled_deduplicated_vi",
+}
+sharded_languages = {
+    "unshuffled_deduplicated_en",
+    "unshuffled_deduplicated_ru",
+    "unshuffled_deduplicated_de",
+    "unshuffled_deduplicated_es",
+    "unshuffled_deduplicated_fr",
+    "unshuffled_deduplicated_ja",
+    "unshuffled_deduplicated_zh",
+}
+
+### Save jsonl
+
+# important: shuffling makes the process 5-7 times slower! best to shuffle the end jsonl file using
+# https://github.com/alexandres/terashuf (should take ~1h to shuffle 900GB file with 70M records
+# using 150GB RAM)
+
+# version 1: one writer - quite slow
+#shuffled_dataset = filtered_dataset.shuffle()
+#shuffled_dataset = filtered_dataset
+#shuffled_dataset.to_json(f"{DATASET_NAME}.jsonl", orient="records", lines=True, force_ascii=False)
+
+# version 2: multiple parallel sharded writes
+# much faster, but will require concatenation at the end
+# 10 shards proved to much for the instance and 3 processed were killed, 5 worked well
+# took about 1.5h per shard
+
+N_SHARDS = 5
+def process_shard(dataset, n_shards, idx, language_subset):
+    if n_shards > 1:
+        print(f"Sharding {idx}")
+        ds_shard = dataset.shard(n_shards, idx, contiguous=True)
+        # shuffle will make things much much slower
+        #ds_shard = ds_shard.shuffle() # remove contiguous=True above if shuffling
+    else:
+        ds_shard = dataset
+    print(f"Saving {DATASET_NAME}-{language_subset}-{idx}.jsonl")
+    export_filename = f"{DATASET_NAME}-{language_subset}-{idx}.jsonl" if n_shards > 1 else \
+        f"{DATASET_NAME}-{language_subset}.jsonl"
+    ds_shard.to_json(export_filename, orient="records", lines=True, force_ascii=False)
+
+for language_subset in language_subsets:
+    dataset = load_dataset(DATASET_NAME, language_subset, split=split, keep_in_memory=False, ignore_verifications=True)
+    n_shards = N_SHARDS if language_subset in sharded_languages else 1
+    queue = Queue()
+    processes = [Process(target=process_shard, args=(dataset, n_shards, idx, language_subset,)) for idx in range(n_shards)]
+    for p in processes:
+        p.start()
+
+    for p in processes:
+        p.join()

data/oscar-multilingual/oscar-to-backup-tgz.slurm

@@ -0,0 +1,26 @@
+#!/bin/bash
+#SBATCH --job-name=oscar-to-backup-tgz # job name
+#SBATCH --ntasks=1                   # number of MP tasks
+#SBATCH --nodes=1
+#SBATCH --cpus-per-task=10           # number of cores per task
+#SBATCH --hint=nomultithread         # we get physical cores not logical
+#SBATCH --time=100:00:00             # maximum execution time (HH:MM:SS)
+#SBATCH --output=%x-%j.out           # output file name
+#SBATCH --partition=cpu_p1
+#SBATCH --account=six@cpu
+#SBATCH --qos=qos_cpu-t4
+
+# 20h is not enough to gzip 1.2TB file, so have to use the other allocation
+##SBATCH --partition=archive
+
+
+set -x -e
+
+cd $six_ALL_CCFRSCRATCH/datasets/oscar-small
+
+# plain text -> gz
+gzip oscar-en-shuffled.jsonl
+mv oscar-en-shuffled.jsonl.gz $six_ALL_CCFRSTORE/datasets/
+
+# already binary -> tar
+tar -cvf $six_ALL_CCFRSTORE/datasets/oscar-en-cache.tar cache

data/p3/prepare_p3.py

@@ -0,0 +1,366 @@
+from functools import partial
+import os
+import multiprocessing
+from datasets import load_dataset, load_from_disk
+import jsonlines
+
+"""Get task list:
+!git clone https://github.com/bigscience-workshop/t-zero.git
+%cd t-zero
+!pip install -e .[seqio_tasks]
+!pip install -q py7zr
+
+import t0.seqio_tasks
+import seqio
+tasks = [task.name for task in seqio.MixtureRegistry.get('t0_train').tasks]
+print(tasks)
+
+
+
+After running the script, merge train & validation jsonls separately into two big files:
+cat folder_with_all_jsonl/*.jsonl > merged_file.jsonl
+"""
+TZERO_TASK_LIST = [
+    'adversarial_qa_dbert_answer_the_following_q',
+    'adversarial_qa_dbert_based_on',
+    'adversarial_qa_dbert_generate_question',
+    'adversarial_qa_dbert_question_context_answer',
+    'adversarial_qa_dbert_tell_what_it_is',
+    'adversarial_qa_dbidaf_answer_the_following_q',
+    'adversarial_qa_dbidaf_based_on',
+    'adversarial_qa_dbidaf_generate_question',
+    'adversarial_qa_dbidaf_question_context_answer',
+    'adversarial_qa_dbidaf_tell_what_it_is',
+    'adversarial_qa_droberta_answer_the_following_q',
+    'adversarial_qa_droberta_based_on',
+    'adversarial_qa_droberta_generate_question',
+    'adversarial_qa_droberta_question_context_answer',
+    'adversarial_qa_droberta_tell_what_it_is',
+    'ag_news_classify',
+    'ag_news_classify_question_first',
+    'ag_news_classify_with_choices',
+    'ag_news_classify_with_choices_question_first',
+    'ag_news_recommend',
+    'ag_news_which_section',
+    'ag_news_which_section_choices',
+    'amazon_polarity_Is_this_product_review_positive',
+    'amazon_polarity_Is_this_review',
+    'amazon_polarity_Is_this_review_negative',
+    'amazon_polarity_User_recommend_this_product',
+    'amazon_polarity_convey_negative_or_positive_sentiment',
+    'amazon_polarity_flattering_or_not',
+    'amazon_polarity_negative_or_positive_tone',
+    'amazon_polarity_user_satisfied',
+    'amazon_polarity_would_you_buy',
+    'app_reviews_categorize_rating_using_review',
+    'app_reviews_convert_to_rating',
+    'app_reviews_convert_to_star_rating',
+    'app_reviews_generate_review',
+    'cnn_dailymail_3.0.0_2_or_3_sentences',
+    'cnn_dailymail_3.0.0_generate_story',
+    'cnn_dailymail_3.0.0_news_card_view',
+    'cnn_dailymail_3.0.0_news_stock',
+    'cnn_dailymail_3.0.0_news_summary',
+    'cnn_dailymail_3.0.0_spice_up_story',
+    'cnn_dailymail_3.0.0_sum_in_brief',
+    'cnn_dailymail_3.0.0_tldr_summary',
+    'cnn_dailymail_3.0.0_write_an_outline',
+    'common_gen_Example_prompt',
+    'common_gen_Given_concepts_type_1',
+    'common_gen_Given_concepts_type_2',
+    'common_gen_Put_together',
+    'common_gen_choice_in_concept_centric_sentence_generation',
+    'common_gen_random_task_template_prompt',
+    'common_gen_sentence_to_concepts',
+    'common_gen_topic_to_sentence',
+    'common_gen_topics_from_the_sentence',
+    'cos_e_v1.11_aligned_with_common_sense',
+    'cos_e_v1.11_description_question_option_id',
+    'cos_e_v1.11_description_question_option_text',
+    'cos_e_v1.11_explain_why_human',
+    'cos_e_v1.11_generate_explanation_given_text',
+    'cos_e_v1.11_i_think',
+    'cos_e_v1.11_question_description_option_id',
+    'cos_e_v1.11_question_description_option_text',
+    'cos_e_v1.11_question_option_description_id',
+    'cos_e_v1.11_question_option_description_text',
+    'cos_e_v1.11_rationale',
+    'cosmos_qa_context_answer_to_question',
+    'cosmos_qa_context_description_question_answer_id',
+    'cosmos_qa_context_description_question_answer_text',
+    'cosmos_qa_context_description_question_text',
+    'cosmos_qa_context_question_description_answer_id',
+    'cosmos_qa_context_question_description_answer_text',
+    'cosmos_qa_context_question_description_text',
+    'cosmos_qa_description_context_question_answer_id',
+    'cosmos_qa_description_context_question_answer_text',
+    'cosmos_qa_description_context_question_text',
+    'cosmos_qa_no_prompt_id',
+    'cosmos_qa_no_prompt_text',
+    'cosmos_qa_only_question_answer',
+    'dbpedia_14_given_a_choice_of_categories_',
+    'dbpedia_14_given_a_list_of_category_what_does_the_title_belong_to',
+    'dbpedia_14_given_list_what_category_does_the_paragraph_belong_to',
+    'dbpedia_14_pick_one_category_for_the_following_text',
+    'dream_answer_to_dialogue',
+    'dream_baseline',
+    'dream_generate_first_utterance',
+    'dream_generate_last_utterance',
+    'dream_read_the_following_conversation_and_answer_the_question',
+    'duorc_ParaphraseRC_answer_question',
+    'duorc_ParaphraseRC_build_story_around_qa',
+    'duorc_ParaphraseRC_decide_worth_it',
+    'duorc_ParaphraseRC_extract_answer',
+    'duorc_ParaphraseRC_generate_question',
+    'duorc_ParaphraseRC_generate_question_by_answer',
+    'duorc_ParaphraseRC_movie_director',
+    'duorc_ParaphraseRC_question_answering',
+    'duorc_ParaphraseRC_title_generation',
+    'duorc_SelfRC_answer_question',
+    'duorc_SelfRC_build_story_around_qa',
+    'duorc_SelfRC_decide_worth_it',
+    'duorc_SelfRC_extract_answer',
+    'duorc_SelfRC_generate_question',
+    'duorc_SelfRC_generate_question_by_answer',
+    'duorc_SelfRC_movie_director',
+    'duorc_SelfRC_question_answering',
+    'duorc_SelfRC_title_generation',
+    'gigaword_TLDR',
+    'gigaword_first_sentence_title',
+    'gigaword_generate_summary_for_this',
+    'gigaword_in_a_nutshell',
+    'gigaword_make_a_title',
+    'gigaword_reverse_writing',
+    'gigaword_write_a_title_for_this_sentence',
+    'gigaword_write_an_article',
+    'gigaword_write_its_sentence',
+    'glue_mrpc_equivalent',
+    'glue_mrpc_generate_paraphrase',
+    'glue_mrpc_generate_sentence',
+    'glue_mrpc_paraphrase',
+    'glue_mrpc_replace',
+    'glue_mrpc_same_thing',
+    'glue_mrpc_want_to_know',
+    'glue_qqp_answer',
+    'glue_qqp_duplicate',
+    'glue_qqp_duplicate_or_not',
+    'glue_qqp_meaning',
+    'glue_qqp_quora',
+    'glue_qqp_same_thing',
+    'imdb_Movie_Expressed_Sentiment',
+    'imdb_Movie_Expressed_Sentiment_2',
+    'imdb_Negation_template_for_positive_and_negative',
+    'imdb_Reviewer_Enjoyment',
+    'imdb_Reviewer_Enjoyment_Yes_No',
+    'imdb_Reviewer_Expressed_Sentiment',
+    'imdb_Reviewer_Opinion_bad_good_choices',
+    'imdb_Reviewer_Sentiment_Feeling',
+    'imdb_Sentiment_with_choices_',
+    'imdb_Text_Expressed_Sentiment',
+    'imdb_Writer_Expressed_Sentiment',
+    'kilt_tasks_hotpotqa_combining_facts',
+    'kilt_tasks_hotpotqa_complex_question',
+    'kilt_tasks_hotpotqa_final_exam',
+    'kilt_tasks_hotpotqa_formulate',
+    'kilt_tasks_hotpotqa_straighforward_qa',
+    'multi_news_distill',
+    'multi_news_expand_reverse_task_',
+    'multi_news_summarize',
+    'multi_news_summary_scenario',
+    'multi_news_synthesize',
+    'multi_news_what_are_the_key_points',
+    'paws_labeled_final_Concatenation',
+    'paws_labeled_final_Concatenation_no_label',
+    'paws_labeled_final_Meaning',
+    'paws_labeled_final_Meaning_no_label',
+    'paws_labeled_final_PAWS_ANLI_GPT3',
+    'paws_labeled_final_PAWS_ANLI_GPT3_no_label',
+    'paws_labeled_final_Rewrite',
+    'paws_labeled_final_Rewrite_no_label',
+    'paws_labeled_final_context_question',
+    'paws_labeled_final_context_question_no_label',
+    'paws_labeled_final_paraphrase_task',
+    'paws_labeled_final_task_description_no_label',
+    'qasc_is_correct_1',
+    'qasc_is_correct_2',
+    'qasc_qa_with_combined_facts_1',
+    'qasc_qa_with_separated_facts_1',
+    'qasc_qa_with_separated_facts_2',
+    'qasc_qa_with_separated_facts_3',
+    'qasc_qa_with_separated_facts_4',
+    'qasc_qa_with_separated_facts_5',
+    'quail_context_description_question_answer_id',
+    'quail_context_description_question_answer_text',
+    'quail_context_description_question_text',
+    'quail_context_question_answer_description_id',
+    'quail_context_question_answer_description_text',
+    'quail_context_question_description_answer_id',
+    'quail_context_question_description_answer_text',
+    'quail_context_question_description_text',
+    'quail_description_context_question_answer_id',
+    'quail_description_context_question_answer_text',
+    'quail_description_context_question_text',
+    'quail_no_prompt_id',
+    'quail_no_prompt_text',
+    'quarel_choose_between',
+    'quarel_do_not_use',
+    'quarel_heres_a_story',
+    'quarel_logic_test',
+    'quarel_testing_students',
+    'quartz_answer_question_based_on',
+    'quartz_answer_question_below',
+    'quartz_given_the_fact_answer_the_q',
+    'quartz_having_read_above_passage',
+    'quartz_paragraph_question_plain_concat',
+    'quartz_read_passage_below_choose',
+    'quartz_use_info_from_paragraph_question',
+    'quartz_use_info_from_question_paragraph',
+    'quoref_Answer_Friend_Question',
+    'quoref_Answer_Question_Given_Context',
+    'quoref_Answer_Test',
+    'quoref_Context_Contains_Answer',
+    'quoref_Find_Answer',
+    'quoref_Found_Context_Online',
+    'quoref_Given_Context_Answer_Question',
+    'quoref_Guess_Answer',
+    'quoref_Guess_Title_For_Context',
+    'quoref_Read_And_Extract_',
+    'quoref_What_Is_The_Answer',
+    'ropes_background_new_situation_answer',
+    'ropes_background_situation_middle',
+    'ropes_given_background_situation',
+    'ropes_new_situation_background_answer',
+    'ropes_plain_background_situation',
+    'ropes_plain_bottom_hint',
+    'ropes_plain_no_background',
+    'ropes_prompt_beginning',
+    'ropes_prompt_bottom_hint_beginning',
+    'ropes_prompt_bottom_no_hint',
+    'ropes_prompt_mix',
+    'ropes_read_background_situation',
+    'rotten_tomatoes_Movie_Expressed_Sentiment',
+    'rotten_tomatoes_Movie_Expressed_Sentiment_2',
+    'rotten_tomatoes_Reviewer_Enjoyment',
+    'rotten_tomatoes_Reviewer_Enjoyment_Yes_No',
+    'rotten_tomatoes_Reviewer_Expressed_Sentiment',
+    'rotten_tomatoes_Reviewer_Opinion_bad_good_choices',
+    'rotten_tomatoes_Reviewer_Sentiment_Feeling',
+    'rotten_tomatoes_Sentiment_with_choices_',
+    'rotten_tomatoes_Text_Expressed_Sentiment',
+    'rotten_tomatoes_Writer_Expressed_Sentiment',
+    'samsum_Generate_a_summary_for_this_dialogue',
+    'samsum_Given_the_above_dialogue_write_a_summary',
+    'samsum_Sum_up_the_following_dialogue',
+    'samsum_Summarize_',
+    'samsum_Summarize_this_dialogue_',
+    'samsum_To_sum_up_this_dialog',
+    'samsum_Write_a_dialogue_that_match_this_summary',
+    'sciq_Direct_Question',
+    'sciq_Direct_Question_Closed_Book_',
+    'sciq_Multiple_Choice',
+    'sciq_Multiple_Choice_Closed_Book_',
+    'sciq_Multiple_Choice_Question_First',
+    'social_i_qa_Check_if_a_random_answer_is_valid_or_not',
+    'social_i_qa_Generate_answer',
+    'social_i_qa_Generate_the_question_from_the_answer',
+    'social_i_qa_I_was_wondering',
+    'social_i_qa_Show_choices_and_generate_answer',
+    'social_i_qa_Show_choices_and_generate_index',
+    'trec_fine_grained_ABBR',
+    'trec_fine_grained_ABBR_context_first',
+    'trec_fine_grained_DESC',
+    'trec_fine_grained_DESC_context_first',
+    'trec_fine_grained_ENTY',
+    'trec_fine_grained_HUM',
+    'trec_fine_grained_HUM_context_first',
+    'trec_fine_grained_LOC',
+    'trec_fine_grained_LOC_context_first',
+    'trec_fine_grained_NUM',
+    'trec_fine_grained_NUM_context_first',
+    'trec_fine_grained_open',
+    'trec_fine_grained_open_context_first',
+    'trec_pick_the_best_descriptor',
+    'trec_trec1',
+    'trec_trec2',
+    'trec_what_category_best_describe',
+    'trec_which_category_best_describes',
+    'wiki_bio_comprehension',
+    'wiki_bio_guess_person',
+    'wiki_bio_key_content',
+    'wiki_bio_what_content',
+    'wiki_bio_who',
+    'wiki_hop_original_choose_best_object_affirmative_1',
+    'wiki_hop_original_choose_best_object_affirmative_2',
+    'wiki_hop_original_choose_best_object_affirmative_3',
+    'wiki_hop_original_choose_best_object_interrogative_1',
+    'wiki_hop_original_choose_best_object_interrogative_2',
+    'wiki_hop_original_explain_relation',
+    'wiki_hop_original_generate_object',
+    'wiki_hop_original_generate_subject',
+    'wiki_hop_original_generate_subject_and_object',
+    'wiki_qa_Decide_good_answer',
+    'wiki_qa_Direct_Answer_to_Question',
+    'wiki_qa_Generate_Question_from_Topic',
+    'wiki_qa_Is_This_True_',
+    'wiki_qa_Jeopardy_style',
+    'wiki_qa_Topic_Prediction_Answer_Only',
+    'wiki_qa_Topic_Prediction_Question_Only',
+    'wiki_qa_Topic_Prediction_Question_and_Answer_Pair',
+    'wiki_qa_automatic_system',
+    'wiki_qa_exercise',
+    'wiki_qa_found_on_google',
+    'wiqa_does_the_supposed_perturbation_have_an_effect',
+    'wiqa_effect_with_label_answer',
+    'wiqa_effect_with_string_answer',
+    'wiqa_what_is_the_final_step_of_the_following_process',
+    'wiqa_what_is_the_missing_first_step',
+    'wiqa_what_might_be_the_first_step_of_the_process',
+    'wiqa_what_might_be_the_last_step_of_the_process',
+    'wiqa_which_of_the_following_is_the_supposed_perturbation',
+    'xsum_DOC_boils_down_to_simple_idea_that',
+    'xsum_DOC_given_above_write_one_sentence',
+    'xsum_DOC_how_would_you_rephrase_few_words',
+    'xsum_DOC_tldr',
+    'xsum_DOC_write_summary_of_above',
+    'xsum_article_DOC_summary',
+    'xsum_college_roommate_asked_DOC_so_I_recap',
+    'xsum_read_below_DOC_write_abstract',
+    'xsum_summarize_DOC',
+    'xsum_summarize_this_DOC_summary',
+    'yelp_review_full_based_on_that',
+    'yelp_review_full_format_rating',
+    'yelp_review_full_format_score',
+    'yelp_review_full_format_star',
+    'yelp_review_full_on_a_scale',
+    'yelp_review_full_so_i_would',
+    'yelp_review_full_this_place'
+]
+
+# Optonally download all first
+# for task_name in TZERO_TASK_LIST:
+#     ds = load_dataset("bigscience/P3", task_name)
+
+def write_to_jsonl_hub(task_name, split):
+    # Could also use ds.to_json()
+    ds = load_dataset("bigscience/P3", task_name)
+    if split in ds:
+        with jsonlines.open(f'p3_{task_name}_{split}.jsonl', mode='w') as writer:
+            for example in ds[split].select(range(len(ds[split]))):
+                writer.write({
+                    "inputs": example["inputs_pretokenized"], 
+                    "targets": example["targets_pretokenized"]
+                })
+
+def write_to_jsonl_disk(task_name, split):
+    ds = load_from_disk(f"{os.environ['six_ALL_CCFRSCRATCH']}/datasets/p3/{task_name}")
+    if split in ds:
+        with jsonlines.open(f'p3_{task_name}_{split}.jsonl', mode='w') as writer:
+            for example in ds[split].select(range(len(ds[split]))):
+                writer.write({
+                    "inputs": example["inputs_pretokenized"], 
+                    "targets": example["targets_pretokenized"]
+                })
+
+with multiprocessing.Pool(num_proc=multiprocessing.cpu_count()) as pool:
+    pool.map(partial(write_to_jsonl_disk, split="train"), TZERO_TASK_LIST)
+    pool.map(partial(write_to_jsonl_disk, split="validation"), TZERO_TASK_LIST)

data/p3/prepare_p3.slurm

@@ -0,0 +1,19 @@
+#!/bin/bash
+#SBATCH --job-name=prepare-p3    # job name
+#SBATCH --ntasks=1                   # number of MP tasks
+#SBATCH --nodes=1
+#SBATCH --cpus-per-task=40         # number of cores per tasks
+#SBATCH --hint=nomultithread         # we get physical cores not logical
+#SBATCH --time=20:00:00             # maximum execution time (HH:MM:SS)
+#SBATCH --output=%x-%j.out           # output file name
+#SBATCH --account=six@cpu
+#SBATCH --partition=compil
+
+set -x -e
+
+source $six_ALL_CCFRWORK/start-prod
+
+# use SCRATCH for building as it's much faster
+cd $six_ALL_CCFRSCRATCH/datasets/bigscience___p3/
+python $six_ALL_CCFRSCRATCH/datasets/bigscience___p3/prepare_p3.py
+

data/sampling_probs/calc_iterator_prob.py

@@ -0,0 +1,132 @@
+import os
+import math
+import json
+import argparse
+import subprocess
+from collections import OrderedDict
+from new_to_old_format_data_path import output_sampling_probs_new_format
+
+SPLIT = [0, 0.949, 0.999, 1.0]
+
+
+def calc_multinomial_sampling_prob_with_penalty(dataset_size, alpha=.5):
+    """
+    Calculate multinomial probability distribution based on https://arxiv.org/pdf/1901.07291.pdf (section 3.1)
+    :dataset_size: A dictionary contains the size (value) of each of the language (key).
+    """
+    tot_size = 0
+    probs = OrderedDict()
+    for lang, size in dataset_size.items():
+        tot_size += size
+    for lang, size in dataset_size.items():
+        probs[lang] = size / tot_size
+
+    pen_prob = OrderedDict()
+    tot_pen_prob = 0.0
+    for lang, prob in probs.items():
+        tot_pen_prob += (prob ** alpha)
+    sum_ = 0.0
+    for lang, prob in probs.items():
+        pen_prob[lang] = (prob ** alpha) / tot_pen_prob
+        sum_ += pen_prob[lang]
+    assert math.fabs(1 - sum_) < 1e-6
+    return pen_prob
+
+
+def get_size_stats(args):
+    """
+    Calculate size for each of the iterator.
+    It recusively iterate though a directory to find a specific extension file and report their size in preferred format.
+    """
+    lang_size_dict = {}
+    for (dirpath, dirnames, filenames) in os.walk(args.data_folder_path):
+        for filename in filenames:
+            if not (filename.startswith(args.name_prefix) and filename.endswith(args.extension_name)):
+                continue
+            full_file_path = os.path.join(dirpath, filename)
+            lang_size = subprocess.check_output("du -s {}".format(full_file_path), shell=True)
+            lang_size = int(lang_size.decode("utf-8").split("\t")[0])
+            if args.size_format == 'KB':
+                _conv = 1
+            elif args.size_format == 'MB':
+                _conv = 1024
+            elif args.size_format == 'GB':
+                _conv = 1024 * 1024
+            elif args.size_format == 'TB':
+                _conv = 1024 * 1024 * 1024
+            lang_size_ = round(lang_size / float(_conv), 2)
+            lang_size_dict[full_file_path] = lang_size_
+    return lang_size_dict
+
+
+def print_stat(args, lang_size_dict, value_name='size'):
+    """
+    Print size statistics.
+    """
+    lang_list = sorted([(k, v) for k, v in lang_size_dict.items()], key=lambda tup: tup[1])
+    total_size = 0
+    print("\nLanguage : ({})".format(value_name))
+    print("-" * 20)
+    for lang, size in lang_list:
+        print("{} :   {}".format(lang, size))
+        total_size += size
+    print("-" * 20)
+    print("Total size : {}".format(total_size))
+
+
+def removesuffix(string, suffix):
+    if string.endswith(suffix):
+        string = string[:-len(suffix)]
+    return string
+
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--data-folder-path', type=str, required=True,
+                        help='Path to the data folder')
+    parser.add_argument('--size-format', type=str, required=True,
+                        help='Calculation will be done in byte, mega-byte, giga-byte or tera-byte',
+                        choices=['KB', 'MB', 'GB', 'TB'])
+    parser.add_argument('--alpha', type=float, required=True,
+                        help='Sampling penalty.')
+    parser.add_argument('--output-dir', type=str, required=True,
+                        help='Output directory where sampling prob_dict will be saved.')
+    parser.add_argument('--name-prefix', type=str, required=True,
+                        help='File name prefix to match. Combination of `--name-prefix` and --extension-name will be used to select file.')
+    parser.add_argument('--extension-name', type=str, required=True,
+                        help='Extension of the file to match. Combination of `--name-prefix` and --extension-name will be used to select file')
+    parser.add_argument('--old-format', action="store_true",
+                        help='Legacy option')
+
+    args = parser.parse_args()
+    size_dict = get_size_stats(args)
+    print_stat(args, size_dict, value_name=args.size_format)
+    sampling_probability = calc_multinomial_sampling_prob_with_penalty(
+        size_dict, alpha=args.alpha
+    )
+    print_stat(args, sampling_probability, 'probability')
+    total_contrib = 0
+    print("\nLanguage : Per epoch contribution in {}".format(args.size_format))
+    print("-" * 50)
+    for lang, prob in sampling_probability.items():
+        sampling_probability[lang] = (prob, size_dict[lang])
+        lang_contrib_size = round(size_dict[lang] * prob, 2)
+        print("{} :   {}  ({}  ->  {})".format(lang, prob, size_dict[lang], lang_contrib_size))
+        total_contrib += lang_contrib_size
+    print("-" * 50)
+    print("Total size : {}".format(total_contrib))
+
+    open(os.path.join(args.output_dir, 'iterator_selection_prob.{}.json'.format(args.alpha)), "w").write(
+        json.dumps(sampling_probability, indent=4)
+    )
+
+    if args.old_format:
+        with open(os.path.join(args.output_dir, "dataset_probabilities.{}.txt".format(args.alpha)), "w") as fout:
+            fout.write(
+                " ".join([f"{prob[0]} {removesuffix(path, '.bin')}" for path, prob in sampling_probability.items()]))
+        pass
+    else:
+        output_sampling_probs_new_format(sampling_probability, args.output_dir, args.alpha)
+
+if __name__ == '__main__':
+    main()

data/sampling_probs/calc_iterator_prob.sh

@@ -0,0 +1,12 @@
+mkdir -p $1
+echo $1
+echo $2
+for alpha in .1 .2 .3 .4 .5 .6 .7 .8 .9; do
+    python data/sampling_probs/calc_iterator_prob.py \
+    --data-folder-path $2/ \
+    --size-format GB \
+    --alpha $alpha \
+    --output-dir $1 \
+    --name-prefix 'train' \
+    --extension-name 'bin'
+done

data/sampling_probs/new_to_old_format_data_path.py

@@ -0,0 +1,70 @@
+import argparse
+import os.path
+from glob import glob
+import json
+
+SPLIT = [0, 0.949, 0.999, 1.0]
+
+
+def finalize_dataset_string(dataset_string):
+    # remove trailing comma in case
+    # surround with quotes
+    if dataset_string.endswith(","):
+        dataset_string = dataset_string[:-1]
+    return '"' + dataset_string + '"'
+
+
+def get_longest_prefix_and_suffix(file1, file2):
+    # we're assuming all filepaths have the same format
+    prefix = max([i for i in range(len(file1)) if file2.startswith(file1[:i])])
+    suffix = min([i for i in range(len(file1) - 1, -1, -1) if file2.endswith(file1[i:])])
+    return prefix, suffix
+
+
+def output_sampling_probs_new_format(sampling_probs, input_dir, alpha):
+    file_weights = [(k[:-4], v[0]) for k, v in sampling_probs.items()]
+
+    prefix, suffix = get_longest_prefix_and_suffix(file_weights[0][0], file_weights[1][0])
+
+    train_split_string = f"{SPLIT[0]}:{SPLIT[1]}"
+    valid_split_string = f"{SPLIT[1]}:{SPLIT[2]}"
+    test_split_string = f"{SPLIT[2]}:{SPLIT[3]}"
+
+    train_string = f"train:"
+    for file, weight in file_weights:
+        train_string += f" {weight} {train_split_string} {file},"
+    train_string = finalize_dataset_string(train_string)
+    with open(os.path.join(input_dir, f"train_data_string.{alpha}.txt"), "w") as f:
+        f.write(train_string)
+
+    valid_strings = ["all_valid:"]
+    for file, weight in file_weights:
+        valid_strings[0] += f" {weight} {valid_split_string} {file},"
+        language_code = file[prefix:suffix]
+        valid_strings.append(f"valid_{language_code}: 1 {valid_split_string} {file}")
+    valid_string = " ".join([finalize_dataset_string(valid_string) for valid_string in valid_strings])
+    with open(os.path.join(input_dir, f"valid_data_string.{alpha}.txt"), "w") as f:
+        f.write(valid_string)
+
+    test_strings = ["all_test:"]
+    for file, weight in file_weights:
+        test_strings[0] += f" {weight} {test_split_string} {file},"
+        language_code = file[prefix:suffix]
+        test_strings.append(f"test_{language_code}: 1 {test_split_string} {file}")
+    test_string = " ".join([finalize_dataset_string(test_string) for test_string in test_strings])
+    with open(os.path.join(input_dir, f"test_data_string.{alpha}.txt"), "w") as f:
+        f.write(test_string)
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--input-files-dir', type=str, required=True,
+                        help='Path to the data folder')
+    args = parser.parse_args()
+
+    for filename in glob(f'{args.input_files_dir}/*.json'):
+        # assuming alpha is of the form 0.x, this could break
+        alpha = filename[-8:-5]
+        # we remove the .bin at the end of the filename
+        sampling_probs = json.load(open(filename))
+        output_sampling_probs_new_format(sampling_probs, args.input_files_dir, alpha)

data/xp3/download_all_datasets.py

@@ -0,0 +1,162 @@
+from multiprocessing import Pool, cpu_count
+import datasets
+from promptsource.utils import load_dataset
+
+all_datasets = [
+    ('glue','mrpc'),
+    ('glue','qqp'),
+    ('paws','labeled_final'),
+    ('ai2_arc','ARC-Challenge'),
+    ('ai2_arc','ARC-Easy'),
+    ('kilt_tasks','hotpotqa'),
+    ('trivia_qa','unfiltered'),
+    ('web_questions',None),
+    ('wiki_qa',None),
+    ('adversarial_qa','dbidaf'),
+    ('adversarial_qa','dbert'),
+    ('adversarial_qa','droberta'),
+    ('duorc','SelfRC'),
+    ('duorc','ParaphraseRC'),
+    ('ropes',None),
+    ('squad_v2',None),
+    ('super_glue','record'),
+    ('quoref',None),
+    ('cos_e','v1.11'),
+    ('cosmos_qa',None),
+    ('dream',None),
+    ('openbookqa','main'),
+    ('qasc',None),
+    ('quail',None),
+    ('quarel',None),
+    ('quartz',None),
+    ('race','high'),
+    ('race','middle'),
+    ('sciq',None),
+    ('social_i_qa',None),
+    ('super_glue','boolq'),
+    ('super_glue','multirc'),
+    ('wiki_hop','original'),
+    ('wiqa',None),
+    ('piqa',None),
+    ('amazon_polarity',None),
+    ('app_reviews',None),
+    ('imdb',None),
+    ('rotten_tomatoes',None),
+    ('yelp_review_full',None),
+    ('common_gen',None),
+    ('wiki_bio',None),
+    ('cnn_dailymail','3.0.0'),
+    ('gigaword',None),
+    ('multi_news',None),
+    ('samsum',None),
+    ('xsum',None),
+    ('ag_news',None),
+    ('dbpedia_14',None),
+    ('trec',None),
+    # Multilingual
+    ('GEM/wiki_lingua', 'ar'),
+    ('GEM/wiki_lingua', 'en'),
+    ('GEM/wiki_lingua', 'es'),
+    ('GEM/wiki_lingua', 'fr'),
+    ('GEM/wiki_lingua', 'hi'),
+    ('GEM/wiki_lingua', 'id'),
+    ('GEM/wiki_lingua', 'pt'),
+    ('GEM/wiki_lingua', 'vi'),
+    ('GEM/wiki_lingua', 'zh'),
+    ('Helsinki-NLP/tatoeba_mt', 'ara-eng'),
+    ('Helsinki-NLP/tatoeba_mt', 'ara-fra'),
+    ('Helsinki-NLP/tatoeba_mt', 'ara-spa'),
+    ('Helsinki-NLP/tatoeba_mt', 'ben-eng'),
+    ('Helsinki-NLP/tatoeba_mt', 'cat-eng'),
+    ('Helsinki-NLP/tatoeba_mt', 'cat-fra'),
+    ('Helsinki-NLP/tatoeba_mt', 'cat-por'),
+    ('Helsinki-NLP/tatoeba_mt', 'cat-spa'),
+    ('Helsinki-NLP/tatoeba_mt', 'eng-cmn_Hans'),
+    ('Helsinki-NLP/tatoeba_mt', 'eng-cmn_Hant'),
+    ('Helsinki-NLP/tatoeba_mt', 'eng-eus'),
+    ('Helsinki-NLP/tatoeba_mt', 'eng-fra'),
+    ('Helsinki-NLP/tatoeba_mt', 'eng-hin'),
+    ('Helsinki-NLP/tatoeba_mt', 'eng-ind'),
+    ('Helsinki-NLP/tatoeba_mt', 'eng-mal'),
+    ('Helsinki-NLP/tatoeba_mt', 'eng-mar'),
+    ('Helsinki-NLP/tatoeba_mt', 'eng-por'),
+    ('Helsinki-NLP/tatoeba_mt', 'eng-run'),
+    ('Helsinki-NLP/tatoeba_mt', 'eng-spa'),
+    ('Helsinki-NLP/tatoeba_mt', 'eng-swa'),
+    ('Helsinki-NLP/tatoeba_mt', 'eng-tam'),
+    ('Helsinki-NLP/tatoeba_mt', 'eng-tel'),
+    ('Helsinki-NLP/tatoeba_mt', 'eng-urd'),
+    ('Helsinki-NLP/tatoeba_mt', 'eng-vie'),
+    ('Helsinki-NLP/tatoeba_mt', 'eng-zho'),
+    ('Helsinki-NLP/tatoeba_mt', 'eus-spa'),
+    ('Helsinki-NLP/tatoeba_mt', 'fra-cmn_Hans'),
+    ('Helsinki-NLP/tatoeba_mt', 'fra-cmn_Hant'),
+    ('Helsinki-NLP/tatoeba_mt', 'fra-ind'),
+    ('Helsinki-NLP/tatoeba_mt', 'fra-por'),
+    ('Helsinki-NLP/tatoeba_mt', 'fra-run'),
+    ('Helsinki-NLP/tatoeba_mt', 'fra-spa'),
+    ('Helsinki-NLP/tatoeba_mt', 'fra-vie'),
+    ('Helsinki-NLP/tatoeba_mt', 'fra-zho'),
+    ('Helsinki-NLP/tatoeba_mt', 'hin-urd'),
+    ('Helsinki-NLP/tatoeba_mt', 'hin-zho'),
+    ('Helsinki-NLP/tatoeba_mt', 'por-cmn_Hans'),
+    ('Helsinki-NLP/tatoeba_mt', 'por-cmn_Hant'),
+    ('Helsinki-NLP/tatoeba_mt', 'por-spa'),
+    ('Helsinki-NLP/tatoeba_mt', 'por-zho'),
+    ('Helsinki-NLP/tatoeba_mt', 'run-spa'),
+    ('Helsinki-NLP/tatoeba_mt', 'spa-cmn_Hans'),
+    ('Helsinki-NLP/tatoeba_mt', 'spa-cmn_Hant'),
+    ('Helsinki-NLP/tatoeba_mt', 'spa-vie'),
+    ('Helsinki-NLP/tatoeba_mt', 'spa-zho'),
+    ('Helsinki-NLP/tatoeba_mt', 'vie-cmn_Hans'),
+    ('Helsinki-NLP/tatoeba_mt', 'vie-zho'),
+    ('xquad', 'xquad.ar'),
+    ('xquad', 'xquad.zh'),
+    ('xquad', 'xquad.vi'),
+    ('xquad', 'xquad.en'),
+    ('xquad', 'xquad.es'),
+    ('xquad', 'xquad.hi'),
+    ('paws-x', 'en'),
+    ('paws-x', 'es'),
+    ('paws-x', 'fr'),
+    ('paws-x', 'zh'),
+    ('khalidalt/tydiqa-primary', 'arabic'),
+    ('khalidalt/tydiqa-primary', 'bengali'),
+    ('khalidalt/tydiqa-primary', 'english'),
+    ('khalidalt/tydiqa-primary', 'indonesian'),
+    ('khalidalt/tydiqa-primary', 'swahili'),
+    ('khalidalt/tydiqa-primary', 'telugu'),
+    ('khalidalt/tydiqa-goldp', 'arabic'),
+    ('khalidalt/tydiqa-goldp', 'bengali'),
+    ('khalidalt/tydiqa-goldp', 'english'),
+    ('khalidalt/tydiqa-goldp', 'indonesian'),
+    ('khalidalt/tydiqa-goldp', 'swahili'),
+    ('khalidalt/tydiqa-goldp', 'telugu'),
+    ('Muennighoff/mbpp', 'sanitized'),
+    ("openai_humaneval", None),
+    ("great_code", None),
+    ("neural_code_search", "evaluation_dataset"),
+    # flores200
+]
+
+print(all_datasets)
+
+def download(names):
+    d_name, conf_name = names
+    try:
+        if d_name == "Helsinki-NLP/tatoeba_mt":
+            # Fixes a bug when loading a ds where only test split exists
+            ds = datasets.load_dataset(d_name, conf_name, download_config=datasets.DownloadConfig(num_proc=1), ignore_verifications=True, revision="842eb26634a9775f504bb2f3f43cd4cc5f9314d8")
+        else:
+            ds = load_dataset(d_name, conf_name, download_config=datasets.DownloadConfig(num_proc=1))
+    except Exception as e:
+        print(f"--- ERROR Dataset {d_name} {conf_name}\n")
+        print(e)
+        return
+
+with Pool(cpu_count()) as pool:
+    _ = pool.map(
+        download,
+        all_datasets,
+    )
+print("ALL DONE")

data/xp3/p3_jsonl_to_meg_bos.slurm

@@ -0,0 +1,66 @@
+#!/bin/bash
+#SBATCH --job-name=xp3jsonl # job name
+#SBATCH --ntasks=1                   # number of MP tasks
+#SBATCH --nodes=1
+#SBATCH --cpus-per-task=40           # number of cores per tasks
+#SBATCH --hint=nomultithread         # we get physical cores not logical
+#SBATCH --time=100:00:00             # maximum execution time (HH:MM:SS)
+#SBATCH --output=%x-%j.out          # output file name
+#SBATCH --account=six@cpu
+#SBATCH --partition=cpu_p1
+
+set -x -e
+
+source $six_ALL_CCFRWORK/start-tr13f-6B3-ml-t0
+export HF_DATASETS_OFFLINE=1
+export TRANSFORMERS_OFFLINE=1
+
+MEGATRON_DEEPSPEED_REPO=$six_ALL_CCFRWORK/code/tr13f-6B3-ml-t0/Megatron-DeepSpeed
+cd $MEGATRON_DEEPSPEED_REPO
+
+
+DATA_PATH=/gpfswork/rech/six/commun/bigscience-training/jsonls/p31t0/p31t0_train.jsonl
+OUTPUT=/gpfswork/rech/six/commun/bigscience-training/p31t0bos/p31t0_train
+TOKENIZER_PATH="bigscience/tokenizer"
+python tools/preprocess_data.py \
+    --input $DATA_PATH \
+    --output-prefix $OUTPUT \
+    --dataset-impl mmap \
+    --json-key inputs \
+    --tokenizer-type PretrainedFromHF \
+    --tokenizer-name-or-path $TOKENIZER_PATH \
+    --append-bos \
+    --workers 35
+python tools/preprocess_data.py \
+    --input $DATA_PATH \
+    --output-prefix $OUTPUT \
+    --dataset-impl mmap \
+    --json-key targets \
+    --tokenizer-type PretrainedFromHF \
+    --tokenizer-name-or-path $TOKENIZER_PATH \
+    --append-eod \
+    --workers 35
+
+
+DATA_PATH=/gpfswork/rech/six/commun/bigscience-training/jsonls/p31t0/p31t0_validation.jsonl
+OUTPUT=/gpfswork/rech/six/commun/bigscience-training/p31t0bos/p31t0_validation
+TOKENIZER_PATH="bigscience/tokenizer"
+
+python tools/preprocess_data.py \
+    --input $DATA_PATH \
+    --output-prefix $OUTPUT \
+    --dataset-impl mmap \
+    --json-key inputs \
+    --tokenizer-type PretrainedFromHF \
+    --tokenizer-name-or-path $TOKENIZER_PATH \
+    --append-bos \
+    --workers 35
+python tools/preprocess_data.py \
+    --input $DATA_PATH \
+    --output-prefix $OUTPUT \
+    --dataset-impl mmap \
+    --json-key targets \
+    --tokenizer-type PretrainedFromHF \
+    --tokenizer-name-or-path $TOKENIZER_PATH \
+    --append-eod \
+    --workers 35

data/xp3/p3_jsonl_to_meg_eos.slurm

@@ -0,0 +1,66 @@
+#!/bin/bash
+#SBATCH --job-name=xp3jsonl # job name
+#SBATCH --ntasks=1                   # number of MP tasks
+#SBATCH --nodes=1
+#SBATCH --cpus-per-task=40           # number of cores per tasks
+#SBATCH --hint=nomultithread         # we get physical cores not logical
+#SBATCH --time=100:00:00             # maximum execution time (HH:MM:SS)
+#SBATCH --output=%x-%j.out          # output file name
+#SBATCH --account=six@cpu
+#SBATCH --partition=cpu_p1
+
+set -x -e
+
+source $six_ALL_CCFRWORK/start-tr13f-6B3-ml-t0
+export HF_DATASETS_OFFLINE=1
+export TRANSFORMERS_OFFLINE=1
+
+MEGATRON_DEEPSPEED_REPO=$six_ALL_CCFRWORK/code/tr13f-6B3-ml-t0/Megatron-DeepSpeed
+cd $MEGATRON_DEEPSPEED_REPO
+
+
+DATA_PATH=/gpfswork/rech/six/commun/bigscience-training/jsonls/p31t0/p31t0_train.jsonl
+OUTPUT=/gpfswork/rech/six/commun/bigscience-training/p31t0eos/p31t0_train
+TOKENIZER_PATH="bigscience/tokenizer"
+python tools/preprocess_data.py \
+    --input $DATA_PATH \
+    --output-prefix $OUTPUT \
+    --dataset-impl mmap \
+    --json-key inputs \
+    --tokenizer-type PretrainedFromHF \
+    --tokenizer-name-or-path $TOKENIZER_PATH \
+    --append-eod \
+    --workers 35
+python tools/preprocess_data.py \
+    --input $DATA_PATH \
+    --output-prefix $OUTPUT \
+    --dataset-impl mmap \
+    --json-key targets \
+    --tokenizer-type PretrainedFromHF \
+    --tokenizer-name-or-path $TOKENIZER_PATH \
+    --append-eod \
+    --workers 35
+
+
+DATA_PATH=/gpfswork/rech/six/commun/bigscience-training/jsonls/p31t0/p31t0_validation.jsonl
+OUTPUT=/gpfswork/rech/six/commun/bigscience-training/p31t0eos/p31t0_validation
+TOKENIZER_PATH="bigscience/tokenizer"
+
+python tools/preprocess_data.py \
+    --input $DATA_PATH \
+    --output-prefix $OUTPUT \
+    --dataset-impl mmap \
+    --json-key inputs \
+    --tokenizer-type PretrainedFromHF \
+    --tokenizer-name-or-path $TOKENIZER_PATH \
+    --append-eod \
+    --workers 35
+python tools/preprocess_data.py \
+    --input $DATA_PATH \
+    --output-prefix $OUTPUT \
+    --dataset-impl mmap \
+    --json-key targets \
+    --tokenizer-type PretrainedFromHF \
+    --tokenizer-name-or-path $TOKENIZER_PATH \
+    --append-eod \
+    --workers 35

data/xp3/prepare_xp3_train.py

@@ -0,0 +1,1194 @@
+from functools import partial
+import json
+import multiprocessing
+import os
+import random
+
+from datasets import load_dataset
+# pip install -q iso-639
+from iso639 import languages
+from promptsource.templates import DatasetTemplates
+
+# Set to False to use multilingual prompts e.g. 'id' for xcopa/id instead of 'en'
+USE_ENGLISH_PROMPTS = True
+
+MAX_EXAMPLES_PER_DATASET_PROMPT = 100_000
+
+STORY_CLOZE_DIR = "/gpfswork/rech/six/commun/code/tr13f-6B3-ml-t0/story_cloze_data"
+XSTORY_CLOZE_DIR = "/gpfswork/rech/six/commun/code/tr13f-6B3-ml-t0/xstory_cloze_data"
+
+# Some datasets have test sets with hidden labels which will still compile but only to noise
+# e.g. piqa test labels are all [-1] which still works on list indices resulting in 
+# noise samples where the label is always the same  
+SKIP_PROMPTS = {
+    "common_gen": {"test": ["all"]},
+    "piqa": {"test": ["all"]},
+    "qasc": {"test": ["all"]},
+    "imdb": {"unsupervised": ["all"]},
+    "glue/qqp": {"test": ["all"]},
+    "qasc": {"test": ["all"]},
+    "cosmos_qa": {"test": [
+        "description_context_question_answer_text", 
+        "description_context_question_text",
+        "description_context_question_answer_id",
+        "context_answer_to_question",
+        "context_description_question_answer_text",
+        "context_description_question_answer_id",
+        "context_question_description_answer_id",
+        "context_description_question_text",
+        "context_question_description_answer_text",
+        "only_question_answer",
+        "no_prompt_id",
+        "context_question_description_text",
+        "no_prompt_text",
+        ]},
+    "clue/tnews": {"test": ["all"]},
+    "clue/csl": {"test": ["all"]},
+    "clue/cmrc2018": {"test": ["generate_question", "in_an_exam", "answer_in_the_passage", "answer_following_question", "xp3longcontinue"]},
+    "clue/drcd": {"test": ["generate_question", "in_an_exam", "answer_in_the_passage", "answer_following_question", "xp3longcontinue"]},
+    "hellaswag": {"test": ["complete_first_then", "Topic of the context", "Open-ended completion", "Randomized prompts template", "Appropriate continuation - Yes or No", "Predict ending with hint", "Open-ended start", "Reversed appropriate continuation - Yes or No", "how_ends", "if_begins_how_continues"]},
+}
+
+DS_TO_ENG_PROMPT = {
+    "xcopa": "en",
+    "Muennighoff/xstory_cloze": "en",
+    "Muennighoff/xwinograd": "en",
+    'GEM/wiki_lingua': 'en_en', # Contains correct language names
+    'xnli': 'en',
+    "paws-x": "en",
+    "mlqa": "mlqa.en.en",
+    "xquad": "xquad.en",
+    "khalidalt/tydiqa-primary": "english",
+    "khalidalt/tydiqa-goldp": "english",
+    "pasinit/xlwic": "en",
+    "GEM/xlsum": "english",
+    "GEM/BiSECT": "en",
+}
+
+BIAS_FAIRNESS = [
+    ('crows_pairs', None),
+    ('jigsaw_toxicity_pred', None),
+    ('super_glue','axg'),
+    ('wino_bias','type1_anti'),
+    ('wino_bias','type2_anti'),
+    ('wino_bias','type1_pro'),
+    ('wino_bias','type2_pro'),
+]
+
+EVAL_DATASETS_L1 = [
+    # ('super_glue','wsc.fixed'), # Not used due to time constraints
+    ('winogrande','winogrande_xl'),
+    ('super_glue','cb'),
+    ('super_glue','rte'),
+    ('anli',None),
+    ('story_cloze', '2016'),
+    ('Muennighoff/xstory_cloze', 'ar'),
+    ('Muennighoff/xstory_cloze', 'es'),
+    ('Muennighoff/xstory_cloze', 'eu'),
+    ('Muennighoff/xstory_cloze', 'id'),
+    ('Muennighoff/xstory_cloze', 'hi'),
+    ('Muennighoff/xstory_cloze', 'te'),
+    ('Muennighoff/xstory_cloze', 'sw'),
+    ('Muennighoff/xstory_cloze', 'zh'),
+    # ('hellaswag', None), # Not used due to time constraints
+    ('super_glue', 'copa'),
+    # Multilingual
+    ('Muennighoff/xwinograd','en'),
+    ('Muennighoff/xwinograd','fr'),
+    ('Muennighoff/xwinograd','pt'),
+    ('Muennighoff/xwinograd','zh'),
+    # ('clue', 'cluewsc2020'), # Included in 'Muennighoff/xwinograd','zh'
+    ('xcopa','id'),
+    ('xcopa','ta'),
+    ('xcopa','sw'),
+    ('xcopa','vi'),
+    ('xcopa','zh'),
+    ("xnli", "ar"),
+    ("xnli", "en"),
+    ("xnli", "es"),
+    ("xnli", "fr"),
+    ("xnli", "hi"),
+    ("xnli", "sw"),
+    ("xnli", "ur"),
+    ("xnli", "vi"),
+    ("xnli", "zh"),
+    # ("openai_humaneval", None), # Used without prompts in evaluation
+    # ("multi_eurlex", "all_languages")
+]
+
+ADD_TRAIN_DATASETS_L1_XP3ALL = [
+    ('super_glue','wsc.fixed'),
+    ('winogrande','winogrande_xl'),
+    ('story_cloze', '2016'),
+    ('Muennighoff/xstory_cloze', 'ar'),
+    ('Muennighoff/xstory_cloze', 'es'),
+    ('Muennighoff/xstory_cloze', 'eu'),
+    ('Muennighoff/xstory_cloze', 'id'),
+    ('Muennighoff/xstory_cloze', 'hi'),
+    ('Muennighoff/xstory_cloze', 'te'),
+    ('Muennighoff/xstory_cloze', 'sw'),
+    ('Muennighoff/xstory_cloze', 'zh'),
+    ('hellaswag', None),
+    ('super_glue', 'copa'),
+    # Multilingual
+    ('Muennighoff/xwinograd','en'),
+    ('Muennighoff/xwinograd','fr'),
+    ('Muennighoff/xwinograd','pt'),
+    ('Muennighoff/xwinograd','zh'),
+    ('clue', 'cluewsc2020'),
+    ('xcopa','id'),
+    ('xcopa','ta'),
+    ('xcopa','sw'),
+    ('xcopa','vi'),
+    ('xcopa','zh'),
+    ("multi_eurlex", "all_languages")
+    # ("openai_humaneval", None), # Low quality prompts
+]
+
+EVAL_DATASETS_L2 = [
+    ('Muennighoff/xwinograd','jp'),
+    ('Muennighoff/xwinograd','ru'),
+    ('xcopa','et'),
+    ('xcopa','ht'),
+    ('xcopa','it'),
+    ('xcopa','qu'),
+    ('xcopa','th'),
+    ('xcopa','tr'),
+    ("xnli", "bg"),
+    ("xnli", "de"),
+    ("xnli", "el"),
+    ("xnli", "ru"),
+    ("xnli", "th"),
+    ("xnli", "tr"),
+]
+
+TRAIN_DATASETS = [
+    # English-only
+    ('glue','mrpc'), 
+    ('glue','qqp'),
+    ('paws','labeled_final'),
+    ('ai2_arc','ARC-Challenge'),
+    ('ai2_arc','ARC-Easy'),
+    ('kilt_tasks','hotpotqa'),
+    ('trivia_qa','unfiltered'),
+    ('web_questions',None),
+    ('wiki_qa',None),
+    ('adversarial_qa','dbidaf'),
+    ('adversarial_qa','dbert'),
+    ('adversarial_qa','droberta'),
+    ('duorc','SelfRC'),
+    ('duorc','ParaphraseRC'),
+    ('ropes',None),
+    ('squad_v2',None),
+    ('super_glue','record'),
+    ('quoref',None),
+    ('cos_e','v1.11'),
+    ('cosmos_qa',None),
+    ('dream',None),
+    ('openbookqa','main'),
+    ('qasc',None),
+    ('quail',None),
+    ('quarel',None),
+    ('quartz',None),
+    ('race','high'),
+    ('race','middle'),
+    ('sciq',None),
+    ('social_i_qa',None),
+    ('super_glue','boolq'),
+    ('super_glue','multirc'),
+    ('wiki_hop','original'),
+    ('wiqa',None),
+    ('piqa',None),
+    ('amazon_polarity',None),
+    ('app_reviews',None),
+    ('imdb',None),
+    ('rotten_tomatoes',None),
+    ('yelp_review_full',None),
+    ('common_gen',None),
+    ('wiki_bio',None),
+    ('cnn_dailymail','3.0.0'),
+    ('gigaword',None),
+    ('multi_news',None),
+    ('samsum',None),
+    ('xsum',None),
+    ('ag_news',None),
+    ('dbpedia_14',None),
+    ('trec',None),
+    # Multilingual
+    ('GEM/wiki_lingua', 'ar'),
+    ('GEM/wiki_lingua', 'en'),
+    ('GEM/wiki_lingua', 'es'),
+    ('GEM/wiki_lingua', 'fr'),
+    ('GEM/wiki_lingua', 'hi'),
+    ('GEM/wiki_lingua', 'id'),
+    ('GEM/wiki_lingua', 'pt'),
+    ('GEM/wiki_lingua', 'vi'),
+    ('GEM/wiki_lingua', 'zh'),
+    ('Helsinki-NLP/tatoeba_mt', 'ara-eng'),
+    ('Helsinki-NLP/tatoeba_mt', 'ara-fra'),
+    ('Helsinki-NLP/tatoeba_mt', 'ara-spa'),
+    ('Helsinki-NLP/tatoeba_mt', 'ben-eng'),
+    ('Helsinki-NLP/tatoeba_mt', 'cat-eng'),
+    ('Helsinki-NLP/tatoeba_mt', 'cat-fra'),
+    ('Helsinki-NLP/tatoeba_mt', 'cat-por'),
+    ('Helsinki-NLP/tatoeba_mt', 'cat-spa'),
+    ('Helsinki-NLP/tatoeba_mt', 'eng-cmn_Hans'),
+    ('Helsinki-NLP/tatoeba_mt', 'eng-cmn_Hant'),
+    ('Helsinki-NLP/tatoeba_mt', 'eng-eus'),
+    ('Helsinki-NLP/tatoeba_mt', 'eng-fra'),
+    ('Helsinki-NLP/tatoeba_mt', 'eng-hin'),
+    ('Helsinki-NLP/tatoeba_mt', 'eng-ind'),
+    ('Helsinki-NLP/tatoeba_mt', 'eng-mal'),
+    ('Helsinki-NLP/tatoeba_mt', 'eng-mar'),
+    ('Helsinki-NLP/tatoeba_mt', 'eng-por'),
+    ('Helsinki-NLP/tatoeba_mt', 'eng-run'),
+    ('Helsinki-NLP/tatoeba_mt', 'eng-spa'),
+    ('Helsinki-NLP/tatoeba_mt', 'eng-swa'),
+    ('Helsinki-NLP/tatoeba_mt', 'eng-tam'),
+    ('Helsinki-NLP/tatoeba_mt', 'eng-tel'),
+    ('Helsinki-NLP/tatoeba_mt', 'eng-urd'),
+    ('Helsinki-NLP/tatoeba_mt', 'eng-vie'),
+    ('Helsinki-NLP/tatoeba_mt', 'eng-zho'),
+    ('Helsinki-NLP/tatoeba_mt', 'eus-spa'),
+    ('Helsinki-NLP/tatoeba_mt', 'fra-cmn_Hans'),
+    ('Helsinki-NLP/tatoeba_mt', 'fra-cmn_Hant'),
+    ('Helsinki-NLP/tatoeba_mt', 'fra-ind'),
+    ('Helsinki-NLP/tatoeba_mt', 'fra-por'),
+    ('Helsinki-NLP/tatoeba_mt', 'fra-run'),
+    ('Helsinki-NLP/tatoeba_mt', 'fra-spa'),
+    ('Helsinki-NLP/tatoeba_mt', 'fra-vie'),
+    ('Helsinki-NLP/tatoeba_mt', 'fra-zho'),
+    ('Helsinki-NLP/tatoeba_mt', 'hin-urd'),
+    ('Helsinki-NLP/tatoeba_mt', 'hin-zho'),
+    ('Helsinki-NLP/tatoeba_mt', 'por-cmn_Hans'),
+    ('Helsinki-NLP/tatoeba_mt', 'por-cmn_Hant'),
+    ('Helsinki-NLP/tatoeba_mt', 'por-spa'),
+    ('Helsinki-NLP/tatoeba_mt', 'por-zho'),
+    ('Helsinki-NLP/tatoeba_mt', 'run-spa'),
+    ('Helsinki-NLP/tatoeba_mt', 'spa-cmn_Hans'),
+    ('Helsinki-NLP/tatoeba_mt', 'spa-cmn_Hant'),
+    ('Helsinki-NLP/tatoeba_mt', 'spa-vie'),
+    ('Helsinki-NLP/tatoeba_mt', 'spa-zho'),
+    ('Helsinki-NLP/tatoeba_mt', 'vie-cmn_Hans'),
+    ('Helsinki-NLP/tatoeba_mt', 'vie-zho'),
+    ('xquad', 'xquad.ar'),
+    ('xquad', 'xquad.zh'),
+    ('xquad', 'xquad.vi'),
+    ('xquad', 'xquad.en'),
+    ('xquad', 'xquad.es'),
+    ('xquad', 'xquad.hi'),
+    ('mlqa', 'mlqa.ar.ar'),
+    ('mlqa', 'mlqa.vi.vi'),
+    ('mlqa', 'mlqa.zh.zh'),
+    ('mlqa', 'mlqa.es.es'),
+    ('mlqa', 'mlqa.en.en'),
+    ('mlqa', 'mlqa.hi.hi'),
+
+    ('mlqa', 'mlqa.ar.vi'),
+    ('mlqa', 'mlqa.ar.zh'),
+    ('mlqa', 'mlqa.ar.es'),
+    ('mlqa', 'mlqa.ar.en'),
+    ('mlqa', 'mlqa.ar.hi'),
+
+    ('mlqa', 'mlqa.vi.ar'),
+    ('mlqa', 'mlqa.vi.zh'),
+    ('mlqa', 'mlqa.vi.es'),
+    ('mlqa', 'mlqa.vi.en'),
+    ('mlqa', 'mlqa.vi.hi'),
+
+    ('mlqa', 'mlqa.zh.ar'),
+    ('mlqa', 'mlqa.zh.vi'),
+    ('mlqa', 'mlqa.zh.es'),
+    ('mlqa', 'mlqa.zh.en'),
+    ('mlqa', 'mlqa.zh.hi'),
+
+    ('mlqa', 'mlqa.es.ar'),
+    ('mlqa', 'mlqa.es.vi'),
+    ('mlqa', 'mlqa.es.zh'),
+    ('mlqa', 'mlqa.es.en'),
+    ('mlqa', 'mlqa.es.hi'),
+
+    ('mlqa', 'mlqa.en.ar'),
+    ('mlqa', 'mlqa.es.vi'),
+    ('mlqa', 'mlqa.es.zh'),
+    ('mlqa', 'mlqa.es.es'),
+    ('mlqa', 'mlqa.es.hi'),
+
+    ('mlqa', 'mlqa.hi.ar'),
+    ('mlqa', 'mlqa.hi.vi'),
+    ('mlqa', 'mlqa.hi.zh'),
+    ('mlqa', 'mlqa.hi.es'),
+    ('mlqa', 'mlqa.hi.en'),
+
+    ('paws-x', 'en'),
+    ('paws-x', 'es'),
+    ('paws-x', 'fr'),
+    ('paws-x', 'zh'),
+    ('khalidalt/tydiqa-primary', 'arabic'),
+    ('khalidalt/tydiqa-primary', 'bengali'),
+    ('khalidalt/tydiqa-primary', 'english'),
+    ('khalidalt/tydiqa-primary', 'indonesian'),
+    ('khalidalt/tydiqa-primary', 'swahili'),
+    ('khalidalt/tydiqa-primary', 'telugu'),
+    ('khalidalt/tydiqa-goldp', 'arabic'),
+    ('khalidalt/tydiqa-goldp', 'bengali'),
+    ('khalidalt/tydiqa-goldp', 'english'),
+    ('khalidalt/tydiqa-goldp', 'indonesian'),
+    ('khalidalt/tydiqa-goldp', 'swahili'),
+    ('khalidalt/tydiqa-goldp', 'telugu'),
+    ('Muennighoff/mbpp', 'sanitized'),
+    ("great_code", None),
+    ("neural_code_search", "evaluation_dataset"),
+    ("codeparrot/codecomplex", "codeparrot--codecomplex"),
+    ("codeparrot/github-jupyter-text-code-pairs", None),
+    ("codeparrot/apps", "all"),
+    ("codeparrot/xlcost-text-to-code", "Python-program-level"),
+    ("codeparrot/xlcost-text-to-code", "C-program-level"),
+    ("codeparrot/xlcost-text-to-code", "C++-program-level"),
+    ("codeparrot/xlcost-text-to-code", "Csharp-program-level"),
+    ("codeparrot/xlcost-text-to-code", "Java-program-level"),
+    ("codeparrot/xlcost-text-to-code", "Javascript-program-level"),
+    ("codeparrot/xlcost-text-to-code", "PHP-program-level"),
+    ("teven/code_contests", None),
+    ("teven/code_docstring_corpus", "top_level"),
+    ("Fraser/python-state-changes", None),
+    ('clue', 'c3'),
+    ('clue', 'cmrc2018'),
+    ('clue', 'csl'),
+    ('clue', 'drcd'),
+    ('clue', 'tnews'),
+    ('super_glue', 'wic'),
+    ('pasinit/xlwic', "xlwic_en_zh"),
+    ('pasinit/xlwic', "xlwic_fr_fr"),
+    ('GEM/BiSECT', "en"),
+    ('GEM/BiSECT', "es"),
+    ('GEM/BiSECT', "fr"),
+    ('GEM/xlsum', "arabic"),
+    ('GEM/xlsum', "bengali"),
+    ('GEM/xlsum', "chinese_simplified"),
+    ('GEM/xlsum', "chinese_traditional"),
+    ('GEM/xlsum', "english"),
+    ('GEM/xlsum', "french"),
+    ('GEM/xlsum', "gujarati"),
+    ('GEM/xlsum', "hindi"),
+    ('GEM/xlsum', "igbo"),
+    ('GEM/xlsum', "indonesian"),
+    ('GEM/xlsum', "kirundi"),
+    ('GEM/xlsum', "marathi"),
+    ('GEM/xlsum', "nepali"),
+    ('GEM/xlsum', "portuguese"),
+    ('GEM/xlsum', "punjabi"),
+    ('GEM/xlsum', "spanish"),
+    ('GEM/xlsum', "swahili"),
+    ('GEM/xlsum', "tamil"),
+    ('GEM/xlsum', "telugu"),
+    ('GEM/xlsum', "urdu"),
+    ('GEM/xlsum', "vietnamese"),
+    ('GEM/xlsum', "yoruba"),
+    # flores200, wmt & more wikilingua added below
+]
+
+FLORES_LANGS = [
+    ("Acehnese (Arabic script)", "ace_Arab"),
+    ("Acehnese (Latin script)", "ace_Latn"),
+    ("Mesopotamian Arabic", "acm_Arab"),
+    ("Ta’izzi-Adeni Arabic", "acq_Arab"),
+    ("Tunisian Arabic", "aeb_Arab"),
+    ("Afrikaans", "afr_Latn"),
+    ("South Levantine Arabic", "ajp_Arab"),
+    ("Akan", "aka_Latn"),
+    ("Amharic", "amh_Ethi"),
+    ("North Levantine Arabic", "apc_Arab"),
+    ("Modern Standard Arabic", "arb_Arab"),
+    ("Modern Standard Arabic (Romanized)", "arb_Latn"),
+    ("Najdi Arabic", "ars_Arab"),
+    ("Moroccan Arabic", "ary_Arab"),
+    ("Egyptian Arabic", "arz_Arab"),
+    ("Assamese", "asm_Beng"),
+    ("Asturian", "ast_Latn"),
+    ("Awadhi", "awa_Deva"),
+    ("Central Aymara", "ayr_Latn"),
+    ("South Azerbaijani", "azb_Arab"),
+    ("North Azerbaijani", "azj_Latn"),
+    ("Bashkir", "bak_Cyrl"),
+    ("Bambara", "bam_Latn"),
+    ("Balinese", "ban_Latn"),
+    ("Belarusian", "bel_Cyrl"),
+    ("Bemba", "bem_Latn"),
+    ("Bengali", "ben_Beng"),
+    ("Bhojpuri", "bho_Deva"),
+    ("Banjar (Arabic script)", "bjn_Arab"),
+    ("Banjar (Latin script)", "bjn_Latn"),
+    ("Standard Tibetan", "bod_Tibt"),
+    ("Bosnian", "bos_Latn"),
+    ("Buginese", "bug_Latn"),
+    ("Bulgarian", "bul_Cyrl"),
+    ("Catalan", "cat_Latn"),
+    ("Cebuano", "ceb_Latn"),
+    ("Czech", "ces_Latn"),
+    ("Chokwe", "cjk_Latn"),
+    ("Central Kurdish", "ckb_Arab"),
+    ("Crimean Tatar", "crh_Latn"),
+    ("Welsh", "cym_Latn"),
+    ("Danish", "dan_Latn"),
+    ("German", "deu_Latn"),
+    ("Southwestern Dinka", "dik_Latn"),
+    ("Dyula", "dyu_Latn"),
+    ("Dzongkha", "dzo_Tibt"),
+    ("Greek", "ell_Grek"),
+    ("English", "eng_Latn"),
+    ("Esperanto", "epo_Latn"),
+    ("Estonian", "est_Latn"),
+    ("Basque", "eus_Latn"),
+    ("Ewe", "ewe_Latn"),
+    ("Faroese", "fao_Latn"),
+    ("Fijian", "fij_Latn"),
+    ("Finnish", "fin_Latn"),
+    ("Fon", "fon_Latn"),
+    ("French", "fra_Latn"),
+    ("Friulian", "fur_Latn"),
+    ("Nigerian Fulfulde", "fuv_Latn"),
+    ("Scottish Gaelic", "gla_Latn"),
+    ("Irish", "gle_Latn"),
+    ("Galician", "glg_Latn"),
+    ("Guarani", "grn_Latn"),
+    ("Gujarati", "guj_Gujr"),
+    ("Haitian Creole", "hat_Latn"),
+    ("Hausa", "hau_Latn"),
+    ("Hebrew", "heb_Hebr"),
+    ("Hindi", "hin_Deva"),
+    ("Chhattisgarhi", "hne_Deva"),
+    ("Croatian", "hrv_Latn"),
+    ("Hungarian", "hun_Latn"),
+    ("Armenian", "hye_Armn"),
+    ("Igbo", "ibo_Latn"),
+    ("Ilocano", "ilo_Latn"),
+    ("Indonesian", "ind_Latn"),
+    ("Icelandic", "isl_Latn"),
+    ("Italian", "ita_Latn"),
+    ("Javanese", "jav_Latn"),
+    ("Japanese", "jpn_Jpan"),
+    ("Kabyle", "kab_Latn"),
+    ("Jingpho", "kac_Latn"),
+    ("Kamba", "kam_Latn"),
+    ("Kannada", "kan_Knda"),
+    ("Kashmiri (Arabic script)", "kas_Arab"),
+    ("Kashmiri (Devanagari script)", "kas_Deva"),
+    ("Georgian", "kat_Geor"),
+    ("Central Kanuri (Arabic script)", "knc_Arab"),
+    ("Central Kanuri (Latin script)", "knc_Latn"),
+    ("Kazakh", "kaz_Cyrl"),
+    ("Kabiyè", "kbp_Latn"),
+    ("Kabuverdianu", "kea_Latn"),
+    ("Khmer", "khm_Khmr"),
+    ("Kikuyu", "kik_Latn"),
+    ("Kinyarwanda", "kin_Latn"),
+    ("Kyrgyz", "kir_Cyrl"),
+    ("Kimbundu", "kmb_Latn"),
+    ("Northern Kurdish", "kmr_Latn"),
+    ("Kikongo", "kon_Latn"),
+    ("Korean", "kor_Hang"),
+    ("Lao", "lao_Laoo"),
+    ("Ligurian", "lij_Latn"),
+    ("Limburgish", "lim_Latn"),
+    ("Lingala", "lin_Latn"),
+    ("Lithuanian", "lit_Latn"),
+    ("Lombard", "lmo_Latn"),
+    ("Latgalian", "ltg_Latn"),
+    ("Luxembourgish", "ltz_Latn"),
+    ("Luba-Kasai", "lua_Latn"),
+    ("Ganda", "lug_Latn"),
+    ("Luo", "luo_Latn"),
+    ("Mizo", "lus_Latn"),
+    ("Standard Latvian", "lvs_Latn"),
+    ("Magahi", "mag_Deva"),
+    ("Maithili", "mai_Deva"),
+    ("Malayalam", "mal_Mlym"),
+    ("Marathi", "mar_Deva"),
+    ("Minangkabau (Arabic script)", "min_Arab"),
+    ("Minangkabau (Latin script)", "min_Latn"),
+    ("Macedonian", "mkd_Cyrl"),
+    ("Plateau Malagasy", "plt_Latn"),
+    ("Maltese", "mlt_Latn"),
+    ("Meitei (Bengali script)", "mni_Beng"),
+    ("Halh Mongolian", "khk_Cyrl"),
+    ("Mossi", "mos_Latn"),
+    ("Maori", "mri_Latn"),
+    ("Burmese", "mya_Mymr"),
+    ("Dutch", "nld_Latn"),
+    ("Norwegian Nynorsk", "nno_Latn"),
+    ("Norwegian Bokmål", "nob_Latn"),
+    ("Nepali", "npi_Deva"),
+    ("Northern Sotho", "nso_Latn"),
+    ("Nuer", "nus_Latn"),
+    ("Nyanja", "nya_Latn"),
+    ("Occitan", "oci_Latn"),
+    ("West Central Oromo", "gaz_Latn"),
+    ("Odia", "ory_Orya"),
+    ("Pangasinan", "pag_Latn"),
+    ("Eastern Panjabi", "pan_Guru"),
+    ("Papiamento", "pap_Latn"),
+    ("Western Persian", "pes_Arab"),
+    ("Polish", "pol_Latn"),
+    ("Portuguese", "por_Latn"),
+    ("Dari", "prs_Arab"),
+    ("Southern Pashto", "pbt_Arab"),
+    ("Ayacucho Quechua", "quy_Latn"),
+    ("Romanian", "ron_Latn"),
+    ("Rundi", "run_Latn"),
+    ("Russian", "rus_Cyrl"),
+    ("Sango", "sag_Latn"),
+    ("Sanskrit", "san_Deva"),
+    ("Santali", "sat_Olck"),
+    ("Sicilian", "scn_Latn"),
+    ("Shan", "shn_Mymr"),
+    ("Sinhala", "sin_Sinh"),
+    ("Slovak", "slk_Latn"),
+    ("Slovenian", "slv_Latn"),
+    ("Samoan", "smo_Latn"),
+    ("Shona", "sna_Latn"),
+    ("Sindhi", "snd_Arab"),
+    ("Somali", "som_Latn"),
+    ("Southern Sotho", "sot_Latn"),
+    ("Spanish", "spa_Latn"),
+    ("Tosk Albanian", "als_Latn"),
+    ("Sardinian", "srd_Latn"),
+    ("Serbian", "srp_Cyrl"),
+    ("Swati", "ssw_Latn"),
+    ("Sundanese", "sun_Latn"),
+    ("Swedish", "swe_Latn"),
+    ("Swahili", "swh_Latn"),
+    ("Silesian", "szl_Latn"),
+    ("Tamil", "tam_Taml"),
+    ("Tatar", "tat_Cyrl"),
+    ("Telugu", "tel_Telu"),
+    ("Tajik", "tgk_Cyrl"),
+    ("Tagalog", "tgl_Latn"),
+    ("Thai", "tha_Thai"),
+    ("Tigrinya", "tir_Ethi"),
+    ("Tamasheq (Latin script)", "taq_Latn"),
+    ("Tamasheq (Tifinagh script)", "taq_Tfng"),
+    ("Tok Pisin", "tpi_Latn"),
+    ("Tswana", "tsn_Latn"),
+    ("Tsonga", "tso_Latn"),
+    ("Turkmen", "tuk_Latn"),
+    ("Tumbuka", "tum_Latn"),
+    ("Turkish", "tur_Latn"),
+    ("Twi", "twi_Latn"),
+    ("Central Atlas Tamazight", "tzm_Tfng"),
+    ("Uyghur", "uig_Arab"),
+    ("Ukrainian", "ukr_Cyrl"),
+    ("Umbundu", "umb_Latn"),
+    ("Urdu", "urd_Arab"),
+    ("Northern Uzbek", "uzn_Latn"),
+    ("Venetian", "vec_Latn"),
+    ("Vietnamese", "vie_Latn"),
+    ("Waray", "war_Latn"),
+    ("Wolof", "wol_Latn"),
+    ("Xhosa", "xho_Latn"),
+    ("Eastern Yiddish", "ydd_Hebr"),
+    ("Yoruba", "yor_Latn"),
+    ("Yue Chinese", "yue_Hant"),
+    ("Chinese (Simplified)", "zho_Hans"),
+    ("Chinese (Traditional)", "zho_Hant"),
+    ("Standard Malay", "zsm_Latn"),
+    ("Zulu", "zul_Latn"),
+]
+
+WMT22_LANGS = [
+    ("afr", "eng"),
+    ("afr", "som"),
+    ("amh", "eng"),
+    ("amh", "fra"),
+    ("amh", "nya"),
+    ("amh", "orm"),
+    ("amh", "sna"),
+    ("amh", "som"),
+    ("amh", "ssw"),
+    ("amh", "swh"),
+    ("amh", "tsn"),
+    ("amh", "tso"),
+    ("amh", "umb"),
+    ("amh", "xho"),
+    ("amh", "yor"),
+    ("amh", "zul"),
+    ("eng", "fuv"),
+    ("eng", "hau"),
+    ("eng", "ibo"),
+    ("eng", "kam"),
+    ("eng", "kin"),
+    ("eng", "lin"),
+    ("eng", "lug"),
+    ("eng", "luo"),
+    ("eng", "nso"),
+    ("eng", "nya"),
+    ("eng", "orm"),
+    ("eng", "sna"),
+    ("eng", "som"),
+    ("eng", "ssw"),
+    ("eng", "swh"),
+    ("eng", "tsn"),
+    ("eng", "tso"),
+    ("eng", "umb"),
+    ("eng", "wol"),
+    ("eng", "xho"),
+    ("eng", "yor"),
+    ("eng", "zul"),
+    ("fra", "hau"),
+    ("fra", "ibo"),
+    ("fra", "kam"),
+    ("fra", "kin"),
+    ("fra", "lin"),
+    ("fra", "lug"),
+    ("fra", "luo"),
+    ("fra", "nso"),
+    ("fra", "nya"),
+    ("fra", "orm"),
+    ("fra", "som"),
+    ("fra", "ssw"),
+    ("fra", "swh"),
+    ("fra", "tsn"),
+    ("fra", "tso"),
+    ("fra", "umb"),
+    ("fra", "wol"),
+    ("fra", "xho"),
+    ("fra", "zul"),
+    ("fuv", "hau"),
+    ("fuv", "ibo"),
+    ("fuv", "kam"),
+    ("fuv", "kin"),
+    ("fuv", "lug"),
+    ("fuv", "luo"),
+    ("fuv", "nso"),
+    ("fuv", "nya"),
+    ("fuv", "orm"),
+    ("fuv", "sna"),
+    ("fuv", "som"),
+    ("fuv", "ssw"),
+    ("fuv", "swh"),
+    ("fuv", "tsn"),
+    ("fuv", "tso"),
+    ("fuv", "umb"),
+    ("fuv", "xho"),
+    ("fuv", "yor"),
+    ("fuv", "zul"),
+    ("hau", "ibo"),
+    ("hau", "kam"),
+    ("hau", "kin"),
+    ("hau", "lug"),
+    ("hau", "luo"),
+    ("hau", "nso"),
+    ("hau", "nya"),
+    ("hau", "orm"),
+    ("hau", "sna"),
+    ("hau", "som"),
+    ("hau", "ssw"),
+    ("hau", "swh"),
+    ("hau", "tsn"),
+    ("hau", "tso"),
+    ("hau", "umb"),
+    ("hau", "xho"),
+    ("hau", "yor"),
+    ("hau", "zul"),
+    ("ibo", "kam"),
+    ("ibo", "kin"),
+    ("ibo", "lug"),
+    ("ibo", "luo"),
+    ("ibo", "nso"),
+    ("ibo", "nya"),
+    ("ibo", "orm"),
+    ("ibo", "sna"),
+    ("ibo", "som"),
+    ("ibo", "ssw"),
+    ("ibo", "swh"),
+    ("ibo", "tsn"),
+    ("ibo", "tso"),
+    ("ibo", "umb"),
+    ("ibo", "xho"),
+    ("ibo", "yor"),
+    ("ibo", "zul"),
+    ("kam", "kin"),
+    ("kam", "lug"),
+    ("kam", "luo"),
+    ("kam", "nso"),
+    ("kam", "nya"),
+    ("kam", "orm"),
+    ("kam", "sna"),
+    ("kam", "som"),
+    ("kam", "ssw"),
+    ("kam", "swh"),
+    ("kam", "tsn"),
+    ("kam", "tso"),
+    ("kam", "umb"),
+    ("kam", "xho"),
+    ("kam", "yor"),
+    ("kam", "zul"),
+    ("kin", "lug"),
+    ("kin", "luo"),
+    ("kin", "nso"),
+    ("kin", "nya"),
+    ("kin", "orm"),
+    ("kin", "sna"),
+    ("kin", "som"),
+    ("kin", "ssw"),
+    ("kin", "swh"),
+    ("kin", "tsn"),
+    ("kin", "tso"),
+    ("kin", "umb"),
+    ("kin", "xho"),
+    ("kin", "yor"),
+    ("kin", "zul"),
+    ("lug", "luo"),
+    ("lug", "nso"),
+    ("lug", "nya"),
+    ("lug", "orm"),
+    ("lug", "sna"),
+    ("lug", "som"),
+    ("lug", "ssw"),
+    ("lug", "swh"),
+    ("lug", "tsn"),
+    ("lug", "tso"),
+    ("lug", "umb"),
+    ("lug", "xho"),
+    ("lug", "yor"),
+    ("lug", "zul"),
+    ("luo", "nso"),
+    ("luo", "nya"),
+    ("luo", "orm"),
+    ("luo", "sna"),
+    ("luo", "som"),
+    ("luo", "ssw"),
+    ("luo", "swh"),
+    ("luo", "tsn"),
+    ("luo", "tso"),
+    ("luo", "umb"),
+    ("luo", "xho"),
+    ("luo", "yor"),
+    ("luo", "zul"),
+    ("nso", "nya"),
+    ("nso", "orm"),
+    ("nso", "sna"),
+    ("nso", "som"),
+    ("nso", "ssw"),
+    ("nso", "swh"),
+    ("nso", "tsn"),
+    ("nso", "tso"),
+    ("nso", "umb"),
+    ("nso", "xho"),
+    ("nso", "yor"),
+    ("nso", "zul"),
+    ("nya", "orm"),
+    ("nya", "sna"),
+    ("nya", "som"),
+    ("nya", "ssw"),
+    ("nya", "swh"),
+    ("nya", "tsn"),
+    ("nya", "tso"),
+    ("nya", "umb"),
+    ("nya", "xho"),
+    ("nya", "yor"),
+    ("nya", "zul"),
+    ("orm", "sna"),
+    ("orm", "som"),
+    ("orm", "ssw"),
+    ("orm", "swh"),
+    ("orm", "tsn"),
+    ("orm", "tso"),
+    ("orm", "umb"),
+    ("orm", "xho"),
+    ("orm", "yor"),
+    ("orm", "zul"),
+    ("sna", "som"),
+    ("sna", "ssw"),
+    ("sna", "swh"),
+    ("sna", "tsn"),
+    ("sna", "tso"),
+    ("sna", "umb"),
+    ("sna", "xho"),
+    ("sna", "yor"),
+    ("sna", "zul"),
+    ("som", "ssw"),
+    ("som", "swh"),
+    ("som", "tsn"),
+    ("som", "tso"),
+    ("som", "umb"),
+    ("som", "wol"),
+    ("som", "xho"),
+    ("som", "yor"),
+    ("som", "zul"),
+    ("ssw", "swh"),
+    ("ssw", "tsn"),
+    ("ssw", "tso"),
+    ("ssw", "umb"),
+    ("ssw", "xho"),
+    ("ssw", "yor"),
+    ("ssw", "zul"),
+    ("swh", "tsn"),
+    ("swh", "tso"),
+    ("swh", "umb"),
+    ("swh", "xho"),
+    ("swh", "yor"),
+    ("swh", "zul"),
+    ("tsn", "tso"),
+    ("tsn", "umb"),
+    ("tsn", "xho"),
+    ("tsn", "yor"),
+    ("tsn", "zul"),
+    ("tso", "umb"),
+    ("tso", "xho"),
+    ("tso", "yor"),
+    ("tso", "zul"),
+    ("umb", "xho"),
+    ("umb", "yor"),
+    ("umb", "zul"),
+    ("xho", "yor"),
+    ("xho", "zul"),
+    ("yor", "zul"),
+]
+
+# Copied from metadata
+BLOOM_LANGS = """
+- ak
+- ar
+- as
+- bm
+- bn
+- ca
+- code
+- en
+- es
+- eu
+- fon
+- fr
+- gu
+- hi
+- id
+- ig
+- ki
+- kn
+- lg
+- ln
+- ml
+- mr
+- ne
+- nso
+- ny
+- or
+- pa
+- pt
+- rn
+- rw
+- sn
+- st
+- sw
+- ta
+- te
+- tn
+- ts
+- tum
+- tw
+- ur
+- vi
+- wo
+- xh
+- yo
+- zh
+- zu
+"""
+
+DS_TO_LANG = {
+    'Muennighoff/mbpp': 'code',
+    'openai_humaneval': 'code',
+    "great_code": "code",
+    "neural_code_search": "code",
+    "codeparrot/codecomplex": "code",
+    "codeparrot/github-jupyter-text-code-pairs": "code",
+    "codeparrot/apps": "code",
+    "Fraser/python-state-changes": "code",
+    "codeparrot/xlcost-text-to-code": "code",
+    "teven/code_contests": "code",
+    "teven/code_docstring_corpus": "code",
+    "clue": "zh",
+    "cmn": "zh", # == zho
+    "npi": "ne", # == npe
+    "ory": "or", # == ori
+    "swh": "sw", # == swa
+    "kirundi": "rn", # == rundi
+    "punjabi": "pa", # == panjabi
+    "chinese_simplified": "zh",
+    "chinese_traditional": "zh",
+}
+
+        
+
+bloom_lang_codes_iso3 = []
+bloom_lang_codes_iso2 = []
+for lang in BLOOM_LANGS.split("\n")[1:-1]:
+    iso2 = lang.replace("- ", "")
+    DS_TO_LANG[iso2] = iso2
+    try:
+        name = languages.get(alpha2=iso2)
+        DS_TO_LANG[name.name.lower()] = iso2
+        # name is e.g. 'swahili (macrolanguage)' also add swahili
+        DS_TO_LANG[name.name.lower().split(" ")[0]] = iso2
+
+        iso3 = name.part3
+        DS_TO_LANG[iso3] = iso2
+    except KeyError:
+        print(f"Could not find iso3 code for {lang}.")
+
+# Add GEM multilingual
+WIKILINGUA_LANGS = ["ar", "en", "es", "fr", "hi", "id", "pt", "vi", "zh"]
+for l1_code in WIKILINGUA_LANGS:
+    for l2_code in WIKILINGUA_LANGS:
+        if l1_code == l2_code:
+            continue
+        TRAIN_DATASETS.append(("GEM/wiki_lingua", f"{l1_code}_{l2_code}"))
+
+# Add flores200
+for (l1_name, l1_code) in FLORES_LANGS:
+    for (l2_name, l2_code) in FLORES_LANGS:
+        if l1_code.split("_")[0] not in DS_TO_LANG or l2_code.split("_")[0] not in DS_TO_LANG:
+            print(f"Skipping as {l1_name} or {l2_name} was not pre-trained on.")
+            continue
+        elif l1_name == l2_name:
+            continue
+        TRAIN_DATASETS.append(("facebook/flores", f"{l1_code}-{l2_code}"))
+
+# Add wmt22
+for (l1_code, l2_code) in WMT22_LANGS:
+    if l1_code not in DS_TO_LANG or l2_code not in DS_TO_LANG:
+        print(f"Skipping as {l1_code} or {l2_code} was not pre-trained on.")
+        continue
+    elif l1_code == l2_code:
+        continue
+    TRAIN_DATASETS.append(("allenai/wmt22_african", f"{l1_code}-{l2_code}"))
+
+
+### DATASET CREATION ###
+
+
+# Copied from promptsource.utils
+def removeHyphen(example):
+    example_clean = {}
+    for key in example.keys():
+        if "-" in key:
+            new_key = key.replace("-", "_")
+            example_clean[new_key] = example[key]
+        else:
+            example_clean[key] = example[key]
+    example = example_clean
+    return example
+
+def apply_template(dataset, template, strip_connection=True):
+    def map_fn(ex):
+        ex = removeHyphen(ex)
+        try:
+            inputs_and_targets = template.apply(
+                ex, 
+                strip_connection=strip_connection,
+                truncate=True,
+            )
+        # Skip ValueError("Prompt did not produce an input and at least one target.")
+        # which happens for some prompts with if else clauses based on inputs producing occasional
+        # empty targets
+        except ValueError:
+            return {"inputs": "", "targets": ""}
+        if len(inputs_and_targets) == 2:
+            # Note that the signature changed in promptsource 
+            # In 0.1.0 template.apply returned two strings; In >0.3.0 it retuns a str & list
+            inputs, targets = inputs_and_targets
+            if len(targets) > 1:
+                # Safer to skip, as could be a bug
+                print(f"Found targets longer than 1. Inputs: {inputs} ; Targets {targets}. Skipping.")
+                return {"inputs": "", "targets": ""}
+            targets = targets[0]
+            return {"inputs": inputs, "targets": targets}
+        # When template results in an empty example, template.apply returns [""]
+        # Also, if the template gets split wrong, len can be > 2
+        # We will filter these out later
+        else:
+            # inputs is a str by default & targets a str
+            return {"inputs": "", "targets": ""}
+
+    def filter_fn(ex):
+        return len(ex["inputs"]) > 0 and len(ex["targets"]) > 0
+
+    original_columns = dataset.column_names
+    dataset = dataset.map(map_fn).filter(filter_fn)
+    # map keeps original columns, remove them
+    return dataset.remove_columns(set(original_columns) - {"inputs", "targets"})
+
+def add_language_name_wikilingua(example):
+    example["source_language_name"] = languages.get(alpha2=example["source_language"]).name
+    example["target_language_name"] = languages.get(alpha2=example["target_language"]).name
+    return example
+
+def filter_l1_l2_wikilingua(example, l1, l2):
+    return example["source_language"] == l1 and example["target_language"] == l2
+
+def filter_empty_solution_apps(example):
+    return bool(example["solutions"])
+
+def add_solution_apps(example):
+    example["solution"] = random.choice(json.loads(example["solutions"]))
+    return example
+
+def clean_code_xlcost(example):
+    clean_lines = []
+    cur_indent = 0
+    for line in example["code"].split("NEW_LINE"):
+        cur_indent += line.count("INDENT")
+        cur_indent -= line.count("DEDENT")
+        line = line.replace("INDENT", "").replace("DEDENT", "")
+        line = line.replace("STRNEWLINE", "\n")
+        line = line.replace("TABSYMBOL", "\t")
+        clean_lines.append("\t" * cur_indent + line.strip())
+    example["code_clean"] = "\n".join(clean_lines)
+    return example
+
+def write_to_jsonl_hub(ds, split="train"):
+
+    ### GET DATASET & LANGUAGE ###
+
+    ds_name, subset_name = ds
+
+    is_wikilingua_cross_lingual = (ds_name == "GEM/wiki_lingua") and ("_") in subset_name
+    
+    lang_dir = DS_TO_LANG.get(ds_name, None)
+    if lang_dir is None:
+        lang_dir = DS_TO_LANG.get(subset_name, "en")
+    if ds_name == "facebook/flores":
+        lang_dir = DS_TO_LANG.get(subset_name.split("-")[-1].split("_")[0])
+    elif is_wikilingua_cross_lingual or ds_name == "pasinit/xlwic":
+        lang_dir = DS_TO_LANG.get(subset_name.split("_")[-1])
+    elif ds_name == "xquad":
+        lang_dir = DS_TO_LANG.get(subset_name.split(".")[1])
+    elif ds_name == "mlqa":
+        # Classify it by the target language for cross-lingual (i.e. what the loss is computed on)
+        lang_dir = DS_TO_LANG.get(subset_name.split(".")[1])
+    os.makedirs(lang_dir, exist_ok=True)
+
+    if ds_name == "Helsinki-NLP/tatoeba_mt":
+        ds = load_dataset(ds_name, subset_name, ignore_verifications=True, revision="49aa20ac768eabc5a106a123549ea58053fc9b40")
+    elif ds_name == "story_cloze":   
+        ds = load_dataset(ds_name, subset_name, data_dir=STORY_CLOZE_DIR)
+    elif ds_name == "Muennighoff/xstory_cloze":
+        ds = load_dataset(ds_name, subset_name, data_dir=XSTORY_CLOZE_DIR)
+    else:
+        ds = load_dataset(ds_name, subset_name)
+
+    if ds_name == "GEM/wiki_lingua":
+        # Add names, e.g. Chinese for zh to use them in the jinja prompts
+        ds = ds.map(add_language_name_wikilingua)
+        if is_wikilingua_cross_lingual:
+            # Keep only L1 -> L2 (L2 -> L1 will be a separate dataset)
+            ds = ds.filter(partial(filter_l1_l2_wikilingua, l1=subset_name.split("_")[0], l2=subset_name.split("_")[1]))
+    elif ds_name == "codeparrot/apps":
+        ds = ds.filter(filter_empty_solution_apps).map(add_solution_apps)
+    elif ds_name == "codeparrot/xlcost-text-to-code":
+        ds = ds.map(clean_code_xlcost)
+
+    ### SELECT SPLITS ###
+
+    dataset_splits = list(ds.keys())
+    if subset_name == "xlwic_en_zh":
+        # Train set is en; val & test are zh
+        dataset_splits.remove("train")
+    elif ds_name == "teven/code_docstring_corpus":
+        # Bad quality split
+        dataset_splits.remove("class_level")
+
+    if split == "validation":
+        if split not in dataset_splits or len(dataset_splits) == 1:
+            print(f"Validation not found for {ds_name}")
+            return
+        dataset_splits = ["validation"]
+    elif split == "train":
+        # Use as much as possible
+        # Would need to remove e.g. test datasets to benchmark same task performance
+        if len(dataset_splits) > 1 and "validation" in dataset_splits:
+            dataset_splits.remove("validation")
+        # WikiLingua
+        if "sampled_validation" in dataset_splits:
+            dataset_splits.remove("sampled_validation")
+        if "sampled_test" in dataset_splits:
+            dataset_splits.remove("sampled_test")
+
+    ### SELECT PROMPTS ###
+
+    if subset_name is None:
+        prompt_dataset_name = ds_name
+    else:
+        subset_name_prompt = subset_name
+        if USE_ENGLISH_PROMPTS and ds_name in DS_TO_ENG_PROMPT:
+            subset_name_prompt = DS_TO_ENG_PROMPT[ds_name]
+        prompt_dataset_name = f"{ds_name}/{subset_name_prompt}"
+
+    prompts = DatasetTemplates(prompt_dataset_name)
+
+    ### PROCESS ###
+
+    for split in dataset_splits:
+        for t_name in prompts.all_template_names:
+            print(f"Running {ds_name}/{subset_name}/{split}/{t_name}")
+            if SKIP_PROMPTS.get(prompt_dataset_name, {}).get(split, False):
+                if ("all" in SKIP_PROMPTS[prompt_dataset_name][split]) or (t_name in SKIP_PROMPTS[prompt_dataset_name][split]):
+                    print(f"Skipping DS: {prompt_dataset_name} Split {split} Prompt {t_name}")
+                    continue
+
+            if ds_name == "Helsinki-NLP/tatoeba_mt":
+                # E.g. translate-this-ara-eng, where eng is the target
+                lang_dir = DS_TO_LANG.get(t_name.split("-")[-1].split("_")[0], "en")
+            elif ds_name in ("allenai/wmt22_african", "multi_eurlex"):
+                # One prompt in multi_eurlex has -source+target appended to the languages
+                lang_dir = DS_TO_LANG.get(t_name.replace("-source+target", "").split("-")[-1])
+
+            out_path = os.path.join(
+                lang_dir, 
+                f'xp3_{ds_name}_{subset_name}_{split}_{t_name}.jsonl'.replace("/", "_").replace(" ", "_")
+            )
+            if os.path.exists(out_path):
+                print("Skipping as exists: ", out_path)
+                continue
+            
+            assert len(ds[split]) > 0, f"Got empty: {ds_name}"
+
+            try:
+                if ds_name == "allenai/wmt22_african":
+                    # Sort by laser score, i.e. by increasing confidence & limit samples due to mediocre quality
+                    ds[split] = ds[split].sort("laser_score", reverse=True)
+                    max_range = min(len(ds[split]), MAX_EXAMPLES_PER_DATASET_PROMPT // 2)
+                else:
+                    # Allow 5x buffer for empty examples
+                    max_range = min(len(ds[split]), MAX_EXAMPLES_PER_DATASET_PROMPT * 5)
+                # Shuffle to avoid using the same subset
+                # Leave \n in-between input & targets for code
+                out_ds = apply_template(
+                    dataset=ds[split].shuffle().select(list(range(max_range))), 
+                    template=prompts[t_name],
+                    strip_connection=False if lang_dir == "code" else True
+                )
+                # Keep X shortest examples
+                max_range = min(len(out_ds), MAX_EXAMPLES_PER_DATASET_PROMPT)
+                out_ds = out_ds.sort("inputs").select(list(range(max_range)))
+            except Exception as e:
+                print(f"Skipping due to {e}. DS: {ds_name}/{subset_name} Template: {t_name}")
+                continue
+            # Do not force ascii to allow chars like é
+            if len(out_ds) > 0:
+                out_ds.to_json(out_path, orient="records", lines=True, force_ascii=False)
+
+# Testing:
+#TRAIN_DATASETS = [
+#    ('xquad', 'xquad.ar'),
+#]
+
+#for ds in TRAIN_DATASETS:
+#    write_to_jsonl_hub(ds, split="train")
+
+with multiprocessing.Pool(processes=multiprocessing.cpu_count()) as pool:
+    pool.map(partial(write_to_jsonl_hub, split="train"), TRAIN_DATASETS)
+    pool.map(partial(write_to_jsonl_hub, split="validation"), TRAIN_DATASETS)
+    #pool.map(partial(write_to_jsonl_hub, split="train"), ADD_TRAIN_DATASETS_L1_XP3ALL)
+    #pool.map(partial(write_to_jsonl_hub, split="validation"), ADD_TRAIN_DATASETS_L1_XP3ALL)

data/xp3/prepare_xp3_train.slurm

@@ -0,0 +1,18 @@
+#!/bin/bash
+#SBATCH --job-name=prepare-xp3    # job name
+#SBATCH --ntasks=1                   # number of MP tasks
+#SBATCH --nodes=1
+#SBATCH --cpus-per-task=40         # number of cores per tasks
+#SBATCH --hint=nomultithread         # we get physical cores not logical
+#SBATCH --time=20:00:00             # maximum execution time (HH:MM:SS)
+#SBATCH --output=%x-%j.out           # output file name
+#SBATCH --account=six@cpu
+#SBATCH --partition=compil
+
+set -x -e
+
+source $six_ALL_CCFRWORK/start-prod
+conda activate thomas_t_zero_evaluation
+
+cd /gpfswork/rech/six/commun/bigscience-training/jsonls/xp3long/train/
+python /gpfswork/rech/six/commun/bigscience-training/jsonls/xp3long/train/prepare_xp3.py

data/xp3/update_jsonls.py

@@ -0,0 +1,21 @@
+import glob
+import json
+import os
+import multiprocessing
+
+jsonl_files = glob.glob("/gpfswork/rech/six/commun/bigscience-training/jsonls/xp3capped/*/*/*.jsonl")
+print(jsonl_files)
+
+#for path in jsonl_files:
+def update_jsonl(path):
+    print(path)
+    with open(path, "r") as jsonl_file, open(path.replace(".jsonl", "tmp.jsonl"), "w") as jsonl_file_out:
+        for line in jsonl_file:
+            data = json.loads(line)
+            data["targets"] = data["targets"][0]
+            jsonl_file_out.write(json.dumps(data) + "\n")
+    os.rename(path.replace(".jsonl", "tmp.jsonl"), path)
+
+
+with multiprocessing.Pool(processes=multiprocessing.cpu_count()-5) as pool:
+    pool.map(update_jsonl, jsonl_files)

data/xp3/xp3_jsonl_to_meg.slurm

@@ -0,0 +1,150 @@
+#!/bin/bash
+#SBATCH --job-name=xp3jsonl # job name
+#SBATCH --ntasks=1                   # number of MP tasks
+#SBATCH --nodes=1
+#SBATCH --cpus-per-task=40           # number of cores per tasks
+#SBATCH --hint=nomultithread         # we get physical cores not logical
+#SBATCH --time=20:00:00             # maximum execution time (HH:MM:SS)
+#SBATCH --output=%x-%j.out          # output file name
+#SBATCH --account=six@cpu
+#SBATCH --partition=cpu_p1
+#SBATCH --qos=qos_cpu-t3
+
+set -x -e
+
+source $six_ALL_CCFRWORK/start-tr13f-6B3-ml-t0
+export HF_DATASETS_OFFLINE=1
+export TRANSFORMERS_OFFLINE=1
+
+MEGATRON_DEEPSPEED_REPO=$six_ALL_CCFRWORK/code/tr13f-6B3-ml-t0/Megatron-DeepSpeed
+
+TOKENIZER_PATH="bigscience/tokenizer"
+
+LANGS=(
+ak
+ar
+as
+bm
+bn
+ca
+code
+en
+es
+eu
+fon
+fr
+gu
+hi
+id
+ig
+ki
+kn
+lg
+ln
+ml
+mr
+ne
+nso
+ny
+or
+pa
+pt
+rn
+rw
+sn
+st
+sw
+ta
+te
+tn
+ts
+tum
+tw
+ur
+vi
+wo
+xh
+yo
+zh
+zu
+)
+
+
+DATA_PATH=/gpfswork/rech/six/commun/bigscience-training/jsonls/xp3capped/train
+
+for val in {0..45}; do
+    LANG=${LANGS[$val]}
+    cd $DATA_PATH/$LANG
+    cat *.jsonl > merged_dups_$LANG.jsonl
+    # Drop duplicates (~1G / 37G for en)
+    sort -u merged_dups_$LANG.jsonl | shuf > merged_$LANG.jsonl
+    OUTPUT=/gpfswork/rech/six/commun/bigscience-training/xp3cappednew/train/xp3_train_$LANG
+    cd $MEGATRON_DEEPSPEED_REPO
+    python tools/preprocess_data.py \
+        --input $DATA_PATH/$LANG/merged_$LANG.jsonl \
+        --output-prefix $OUTPUT \
+        --dataset-impl mmap \
+        --json-key inputs \
+        --tokenizer-type PretrainedFromHF \
+        --tokenizer-name-or-path $TOKENIZER_PATH \
+        --workers 35
+    python tools/preprocess_data.py \
+        --input $DATA_PATH/$LANG/merged_$LANG.jsonl \
+        --output-prefix $OUTPUT \
+        --dataset-impl mmap \
+        --json-key targets \
+        --tokenizer-type PretrainedFromHF \
+        --tokenizer-name-or-path $TOKENIZER_PATH \
+        --append-eod \
+        --prepend-space \
+        --workers 35
+done
+
+# No val data for other langs
+LANGS=(
+ar
+bn
+code
+en
+es
+fr
+hi
+id
+pt
+sw
+te
+vi
+zh
+)
+
+DATA_PATH=/gpfswork/rech/six/commun/bigscience-training/jsonls/xp3capped/validation
+cd $DATA_PATH
+
+
+for val in {0..12}; do
+    LANG=${LANGS[$val]}
+    cd $DATA_PATH/$LANG
+    cat *.jsonl > merged_dups_$LANG.jsonl
+    # Drop duplicates (~1G / 37G for en)
+    sort -u merged_dups_$LANG.jsonl > merged_$LANG.jsonl
+    OUTPUT=/gpfswork/rech/six/commun/bigscience-training/xp3cappednew/validation/xp3_validation_$LANG
+    cd $MEGATRON_DEEPSPEED_REPO
+    python tools/preprocess_data.py \
+        --input $DATA_PATH/$LANG/merged_$LANG.jsonl \
+        --output-prefix $OUTPUT \
+        --dataset-impl mmap \
+        --json-key inputs \
+        --tokenizer-type PretrainedFromHF \
+        --tokenizer-name-or-path $TOKENIZER_PATH \
+        --workers 35
+    python tools/preprocess_data.py \
+        --input $DATA_PATH/$LANG/merged_$LANG.jsonl \
+        --output-prefix $OUTPUT \
+        --dataset-impl mmap \
+        --json-key targets \
+        --tokenizer-type PretrainedFromHF \
+        --tokenizer-name-or-path $TOKENIZER_PATH \
+        --append-eod \
+        --prepend-space \
+        --workers 35
+done

data/xp3/xp3cappedmixed_jsonl_to_meg.slurm

@@ -0,0 +1,104 @@
+#!/bin/bash
+#SBATCH --job-name=xp3mixedjsonl # job name
+#SBATCH --ntasks=1                   # number of MP tasks
+#SBATCH --nodes=1
+#SBATCH --cpus-per-task=40           # number of cores per tasks
+#SBATCH --hint=nomultithread         # we get physical cores not logical
+#SBATCH --time=20:00:00             # maximum execution time (HH:MM:SS)
+#SBATCH --output=%x-%j.out          # output file name
+#SBATCH --account=six@cpu
+#SBATCH --partition=cpu_p1
+#SBATCH --qos=qos_cpu-t3
+
+set -x -e
+
+source $six_ALL_CCFRWORK/start-tr13f-6B3-ml-t0
+export HF_DATASETS_OFFLINE=1
+export TRANSFORMERS_OFFLINE=1
+
+MEGATRON_DEEPSPEED_REPO=$six_ALL_CCFRWORK/code/tr13f-6B3-ml-t0/Megatron-DeepSpeed
+
+TOKENIZER_PATH="bigscience/tokenizer"
+
+LANGS=(
+ak
+ar
+as
+bm
+bn
+ca
+code
+en
+es
+eu
+fon
+fr
+gu
+hi
+id
+ig
+ki
+kn
+lg
+ln
+ml
+mr
+ne
+nso
+ny
+or
+pa
+pt
+rn
+rw
+sn
+st
+sw
+ta
+te
+tn
+ts
+tum
+tw
+ur
+vi
+wo
+xh
+yo
+zh
+zu
+)
+
+
+DATA_PATH=/gpfswork/rech/six/commun/bigscience-training/jsonls/xp3cappedmixedfixlong
+OUTPUT=/gpfswork/rech/six/commun/bigscience-training/xp3cappedmixedfixlong
+
+mkdir -p $OUTPUT
+
+for val in {0..45}; do
+    LANG=${LANGS[$val]}
+    cd $DATA_PATH/$LANG
+    # Merge
+    cat *.jsonl > merged_dups_$LANG.jsonl
+    # Drop duplicates (~1G / 37G for en) + Shuffle
+    sort -u merged_dups_$LANG.jsonl | shuf > merged_$LANG.jsonl
+    cd $MEGATRON_DEEPSPEED_REPO
+    python tools/preprocess_data.py \
+        --input $DATA_PATH/$LANG/merged_$LANG.jsonl \
+        --output-prefix $OUTPUT/xp3_$LANG \
+        --dataset-impl mmap \
+        --json-key inputs \
+        --tokenizer-type PretrainedFromHF \
+        --tokenizer-name-or-path $TOKENIZER_PATH \
+        --workers 35
+    python tools/preprocess_data.py \
+        --input $DATA_PATH/$LANG/merged_$LANG.jsonl \
+        --output-prefix $OUTPUT/xp3_$LANG \
+        --dataset-impl mmap \
+        --json-key targets \
+        --tokenizer-type PretrainedFromHF \
+        --tokenizer-name-or-path $TOKENIZER_PATH \
+        --append-eod \
+        --prepend-space \
+        --workers 35
+done

data/xp3/xp3mixed_jsonl_to_meg.slurm

@@ -0,0 +1,102 @@
+#!/bin/bash
+#SBATCH --job-name=xp3mixedjsonl # job name
+#SBATCH --ntasks=1                   # number of MP tasks
+#SBATCH --nodes=1
+#SBATCH --cpus-per-task=40           # number of cores per tasks
+#SBATCH --hint=nomultithread         # we get physical cores not logical
+#SBATCH --time=20:00:00             # maximum execution time (HH:MM:SS)
+#SBATCH --output=%x-%j.out          # output file name
+#SBATCH --account=six@cpu
+#SBATCH --partition=cpu_p1
+#SBATCH --qos=qos_cpu-t3
+
+set -x -e
+
+source $six_ALL_CCFRWORK/start-tr13f-6B3-ml-t0
+export HF_DATASETS_OFFLINE=1
+export TRANSFORMERS_OFFLINE=1
+
+MEGATRON_DEEPSPEED_REPO=$six_ALL_CCFRWORK/code/tr13f-6B3-ml-t0/Megatron-DeepSpeed
+
+TOKENIZER_PATH="bigscience/tokenizer"
+
+LANGS=(
+ak
+ar
+as
+bm
+bn
+ca
+code
+en
+es
+eu
+fon
+fr
+gu
+hi
+id
+ig
+ki
+kn
+lg
+ln
+ml
+mr
+ne
+nso
+ny
+or
+pa
+pt
+rn
+rw
+sn
+st
+sw
+ta
+te
+tn
+ts
+tum
+tw
+ur
+vi
+wo
+xh
+yo
+zh
+zu
+)
+
+
+DATA_PATH=/gpfswork/rech/six/commun/bigscience-training/jsonls/xp3mixed
+
+for val in {0..45}; do
+    LANG=${LANGS[$val]}
+    cd $DATA_PATH/$LANG
+    # Merge
+    cat *.jsonl > merged_dups_$LANG.jsonl
+    # Drop duplicates (~1G / 37G for en) + Shuffle
+    sort -u merged_dups_$LANG.jsonl | shuf > merged_$LANG.jsonl
+    OUTPUT=/gpfswork/rech/six/commun/bigscience-training/xp3mixed/xp3_$LANG
+    cd $MEGATRON_DEEPSPEED_REPO
+    python tools/preprocess_data.py \
+        --input $DATA_PATH/$LANG/merged_$LANG.jsonl \
+        --output-prefix $OUTPUT \
+        --dataset-impl mmap \
+        --json-key inputs \
+        --tokenizer-type PretrainedFromHF \
+        --tokenizer-name-or-path $TOKENIZER_PATH \
+        --workers 35
+    python tools/preprocess_data.py \
+        --input $DATA_PATH/$LANG/merged_$LANG.jsonl \
+        --output-prefix $OUTPUT \
+        --dataset-impl mmap \
+        --json-key targets \
+        --tokenizer-type PretrainedFromHF \
+        --tokenizer-name-or-path $TOKENIZER_PATH \
+        --append-eod \
+        --prepend-space \
+        --workers 35
+done

inference/README.md

@@ -0,0 +1,15 @@
+# Inference
+
+Notes on the plans to do inference with the pre-trained model
+
+# Large Model on limited hardware
+
+- inferencing and tinkering on a single host (150-200B model)
+
+Solution: We can do this with ZeRO-Infinity. Seems like @Shaden Smith already has the code to load the model parameters checkpoints from Megatron+DeepSpeed 3D to Megatron+ DeepSpeed ZeRO-Infinity. The remaining work is to add an inference only mode to ZeRO-Infinity that drops all the non-parameter states.
+
+Hardware Requirements : Would require about 500-1000 GB of memory (can be CPU, GPU or NVMe). Single Node with enough CPU or NVMe memory should work here.
+
+The single node can be as little as 4x 32GB-V100. It will be just slower than say, 8x 80GB-A100.
+
+Estimated Work: If all works as expected, 1-3 weeks based on bandwidth availability. Tuning for the best performance might another week or so, but that wont be blocking the availability of the functionality.

inference/modeling_gpt2_alibi_prefix_lm.py

@@ -0,0 +1,1750 @@
+# coding=utf-8
+# Copyright 2018 The OpenAI Team Authors and HuggingFace Inc. team.
+# Copyright (c) 2018, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""PyTorch OpenAI GPT-2 model with AliBi."""
+
+## integrating some AliBi code from https://github.com/bigscience-workshop/Megatron-DeepSpeed/blob/c839a8aa30731f71b3738d56009be9668508e366/megatron/model/transformer.py
+# I am keeping the name of the classes as GPT2 because some of transformer's code like pipeline classes check class names in order to do things, and
+# creating a new class that have different names sometimes break things.
+
+import os
+import enum
+from dataclasses import dataclass
+from typing import Optional, Tuple
+
+import torch
+import torch.utils.checkpoint
+from torch import nn
+from torch.nn import CrossEntropyLoss, MSELoss
+
+from transformers.activations import ACT2FN
+from transformers.file_utils import (
+    ModelOutput,
+    add_code_sample_docstrings,
+    add_start_docstrings,
+    add_start_docstrings_to_model_forward,
+    replace_return_docstrings,
+)
+from transformers.modeling_outputs import (
+    BaseModelOutputWithPastAndCrossAttentions,
+    CausalLMOutputWithCrossAttentions,
+    SequenceClassifierOutputWithPast,
+    TokenClassifierOutput,
+)
+from transformers.modeling_utils import (
+    Conv1D,
+    PreTrainedModel,
+    SequenceSummary,
+    find_pruneable_heads_and_indices,
+    prune_conv1d_layer,
+)
+from transformers.utils import logging
+from transformers.utils.model_parallel_utils import assert_device_map, get_device_map
+from transformers.models.gpt2.configuration_gpt2 import GPT2Config
+
+from collections import OrderedDict
+from typing import Any, Mapping, Optional
+
+from transformers import PreTrainedTokenizer, TensorType, is_torch_available
+
+from transformers.configuration_utils import PretrainedConfig
+from transformers.onnx import OnnxConfigWithPast
+
+
+
+logger = logging.get_logger(__name__)
+
+GPT2_PRETRAINED_CONFIG_ARCHIVE_MAP = {
+    "gpt2": "https://huggingface.co/gpt2/resolve/main/config.json",
+    "gpt2-medium": "https://huggingface.co/gpt2-medium/resolve/main/config.json",
+    "gpt2-large": "https://huggingface.co/gpt2-large/resolve/main/config.json",
+    "gpt2-xl": "https://huggingface.co/gpt2-xl/resolve/main/config.json",
+    "distilgpt2": "https://huggingface.co/distilgpt2/resolve/main/config.json",
+}
+
+PositionEmbeddingType_rotary = 1 # not implemented
+PositionEmbeddingType_absolute = 2
+PositionEmbeddingType_alibi = 3
+
+
+class GPT2Config(PretrainedConfig):
+    """
+    This is the configuration class to store the configuration of a :class:`~transformers.GPT2Model` or a
+    :class:`~transformers.TFGPT2Model`. It is used to instantiate a GPT-2 model according to the specified arguments,
+    defining the model architecture. Instantiating a configuration with the defaults will yield a similar configuration
+    to that of the GPT-2 `small <https://huggingface.co/gpt2>`__ architecture.
+    Configuration objects inherit from :class:`~transformers.PretrainedConfig` and can be used to control the model
+    outputs. Read the documentation from :class:`~transformers.PretrainedConfig` for more information.
+    Args:
+        vocab_size (:obj:`int`, `optional`, defaults to 50257):
+            Vocabulary size of the GPT-2 model. Defines the number of different tokens that can be represented by the
+            :obj:`inputs_ids` passed when calling :class:`~transformers.GPT2Model` or
+            :class:`~transformers.TFGPT2Model`.
+        n_positions (:obj:`int`, `optional`, defaults to 1024):
+            The maximum sequence length that this model might ever be used with. Typically set this to something large
+            just in case (e.g., 512 or 1024 or 2048).
+        n_ctx (:obj:`int`, `optional`, defaults to 1024):
+            Dimensionality of the causal mask (usually same as n_positions).
+        n_embd (:obj:`int`, `optional`, defaults to 768):
+            Dimensionality of the embeddings and hidden states.
+        n_layer (:obj:`int`, `optional`, defaults to 12):
+            Number of hidden layers in the Transformer encoder.
+        n_head (:obj:`int`, `optional`, defaults to 12):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        n_inner (:obj:`int`, `optional`, defaults to None):
+            Dimensionality of the inner feed-forward layers. :obj:`None` will set it to 4 times n_embd
+        activation_function (:obj:`str`, `optional`, defaults to :obj:`"gelu"`):
+            Activation function, to be selected in the list :obj:`["relu", "silu", "gelu", "tanh", "gelu_new"]`.
+        resid_pdrop (:obj:`float`, `optional`, defaults to 0.1):
+            The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.
+        embd_pdrop (:obj:`int`, `optional`, defaults to 0.1):
+            The dropout ratio for the embeddings.
+        attn_pdrop (:obj:`float`, `optional`, defaults to 0.1):
+            The dropout ratio for the attention.
+        layer_norm_epsilon (:obj:`float`, `optional`, defaults to 1e-5):
+            The epsilon to use in the layer normalization layers
+        initializer_range (:obj:`float`, `optional`, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        summary_type (:obj:`string`, `optional`, defaults to :obj:`"cls_index"`):
+            Argument used when doing sequence summary, used in the models :class:`~transformers.GPT2DoubleHeadsModel`
+            and :class:`~transformers.TFGPT2DoubleHeadsModel`.
+            Has to be one of the following options:
+                - :obj:`"last"`: Take the last token hidden state (like XLNet).
+                - :obj:`"first"`: Take the first token hidden state (like BERT).
+                - :obj:`"mean"`: Take the mean of all tokens hidden states.
+                - :obj:`"cls_index"`: Supply a Tensor of classification token position (like GPT/GPT-2).
+                - :obj:`"attn"`: Not implemented now, use multi-head attention.
+        summary_use_proj (:obj:`bool`, `optional`, defaults to :obj:`True`):
+            Argument used when doing sequence summary, used in the models :class:`~transformers.GPT2DoubleHeadsModel`
+            and :class:`~transformers.TFGPT2DoubleHeadsModel`.
+            Whether or not to add a projection after the vector extraction.
+        summary_activation (:obj:`str`, `optional`):
+            Argument used when doing sequence summary. Used in for the multiple choice head in
+            :class:`~transformers.GPT2DoubleHeadsModel`.
+            Pass :obj:`"tanh"` for a tanh activation to the output, any other value will result in no activation.
+        summary_proj_to_labels (:obj:`bool`, `optional`, defaults to :obj:`True`):
+            Argument used when doing sequence summary, used in the models :class:`~transformers.GPT2DoubleHeadsModel`
+            and :class:`~transformers.TFGPT2DoubleHeadsModel`.
+            Whether the projection outputs should have :obj:`config.num_labels` or :obj:`config.hidden_size` classes.
+        summary_first_dropout (:obj:`float`, `optional`, defaults to 0.1):
+            Argument used when doing sequence summary, used in the models :class:`~transformers.GPT2DoubleHeadsModel`
+            and :class:`~transformers.TFGPT2DoubleHeadsModel`.
+            The dropout ratio to be used after the projection and activation.
+        scale_attn_weights (:obj:`bool`, `optional`, defaults to :obj:`True`):
+            Scale attention weights by dividing by sqrt(hidden_size)..
+        use_cache (:obj:`bool`, `optional`, defaults to :obj:`True`):
+            Whether or not the model should return the last key/values attentions (not used by all models).
+    Example::
+        >>> from transformers import GPT2Model, GPT2Config
+        >>> # Initializing a GPT2 configuration
+        >>> configuration = GPT2Config()
+        >>> # Initializing a model from the configuration
+        >>> model = GPT2Model(configuration)
+        >>> # Accessing the model configuration
+        >>> configuration = model.config
+    """
+
+    model_type = "gpt2"
+    keys_to_ignore_at_inference = ["past_key_values"]
+    attribute_map = {
+        "hidden_size": "n_embd",
+        "max_position_embeddings": "n_positions",
+        "num_attention_heads": "n_head",
+        "num_hidden_layers": "n_layer",
+    }
+
+    def __init__(
+        self,
+        vocab_size=50257,
+        n_positions=1024,
+        n_ctx=1024,
+        n_embd=768,
+        n_layer=12,
+        n_head=12,
+        n_inner=None,
+        activation_function="gelu_new",
+        resid_pdrop=0.1,
+        embd_pdrop=0.1,
+        attn_pdrop=0.1,
+        layer_norm_epsilon=1e-5,
+        initializer_range=0.02,
+        summary_type="cls_index",
+        summary_use_proj=True,
+        summary_activation=None,
+        summary_proj_to_labels=True,
+        summary_first_dropout=0.1,
+        scale_attn_weights=True,
+        use_cache=True,
+        bos_token_id=50256,
+        eos_token_id=50256,
+        position_embedding_type=PositionEmbeddingType_absolute,
+        **kwargs
+    ):
+        self.vocab_size = vocab_size
+        self.n_ctx = n_ctx
+        self.n_positions = n_positions
+        self.n_embd = n_embd
+        self.n_layer = n_layer
+        self.n_head = n_head
+        self.n_inner = n_inner
+        self.activation_function = activation_function
+        self.resid_pdrop = resid_pdrop
+        self.embd_pdrop = embd_pdrop
+        self.attn_pdrop = attn_pdrop
+        self.layer_norm_epsilon = layer_norm_epsilon
+        self.initializer_range = initializer_range
+        self.summary_type = summary_type
+        self.summary_use_proj = summary_use_proj
+        self.summary_activation = summary_activation
+        self.summary_first_dropout = summary_first_dropout
+        self.summary_proj_to_labels = summary_proj_to_labels
+        self.scale_attn_weights = scale_attn_weights
+        self.use_cache = use_cache
+
+        self.bos_token_id = bos_token_id
+        self.eos_token_id = eos_token_id
+        self.position_embedding_type = position_embedding_type
+        super().__init__(bos_token_id=bos_token_id, eos_token_id=eos_token_id, **kwargs)
+
+
+class GPT2OnnxConfig(OnnxConfigWithPast):
+    @property
+    def inputs(self) -> Mapping[str, Mapping[int, str]]:
+        common_inputs = OrderedDict({"input_ids": {0: "batch"}})
+        if self.use_past:
+            for i in range(self._config.n_layer * 2):
+                common_inputs[f"past_key_values.{i}"] = {0: "batch", 2: "sequence"}
+
+            common_inputs["attention_mask"] = {0: "batch", 1: "sequence"}
+        else:
+            common_inputs["attention_mask"] = {0: "batch", 1: "sequence"}
+
+        return common_inputs
+
+    @property
+    def outputs(self) -> Mapping[str, Mapping[int, str]]:
+        common_outputs = OrderedDict({"last_hidden_state": {0: "batch", 1: "sequence"}})
+        if self.use_past:
+            for i in range(self._config.n_layer * 2):
+                common_outputs[f"present.{i}"] = {0: "batch", 2: "sequence"}
+
+            return common_outputs
+
+        return common_outputs
+
+    def generate_dummy_inputs(
+        self,
+        tokenizer: PreTrainedTokenizer,
+        batch_size: int = -1,
+        seq_length: int = -1,
+        is_pair: bool = False,
+        framework: Optional[TensorType] = None,
+    ) -> Mapping[str, Any]:
+        common_inputs = super().generate_dummy_inputs(tokenizer, batch_size, seq_length, is_pair, framework)
+
+        # We need to order the input in the way they appears in the forward()
+        ordered_inputs = OrderedDict({"input_ids": common_inputs["input_ids"]})
+
+        # Need to add the past_keys
+        if self.use_past:
+            if not is_torch_available():
+                raise ValueError("Cannot generate dummy past_keys inputs without PyTorch installed.")
+            else:
+                import torch
+
+                batch = common_inputs["input_ids"].shape[0]
+                ordered_inputs["past_key_values"] = [
+                    (
+                        torch.zeros((batch, self._config.n_head, 1, self._config.hidden_size // self._config.n_head)),
+                        torch.zeros((batch, self._config.n_head, 1, self._config.hidden_size // self._config.n_head)),
+                    )
+                    for _ in range(self._config.n_layer)
+                ]
+
+        ordered_inputs["attention_mask"] = common_inputs["attention_mask"]
+        return ordered_inputs
+
+
+# need to change the checkpoints to be the bigscience checkpoints
+_CHECKPOINT_FOR_DOC = "gpt2"
+_CONFIG_FOR_DOC = "GPT2Config"
+_TOKENIZER_FOR_DOC = "GPT2Tokenizer"
+
+GPT2_PRETRAINED_MODEL_ARCHIVE_LIST = [
+    "gpt2",
+    "gpt2-medium",
+    "gpt2-large",
+    "gpt2-xl",
+    "distilgpt2",
+    # See all GPT-2 models at https://huggingface.co/models?filter=gpt2
+]
+
+
+
+
+def load_tf_weights_in_gpt2(model, config, gpt2_checkpoint_path):
+    """Load tf checkpoints in a pytorch model"""
+    try:
+        import re
+
+        import tensorflow as tf
+    except ImportError:
+        logger.error(
+            "Loading a TensorFlow model in PyTorch, requires TensorFlow to be installed. Please see "
+            "https://www.tensorflow.org/install/ for installation instructions."
+        )
+        raise
+    tf_path = os.path.abspath(gpt2_checkpoint_path)
+    logger.info(f"Converting TensorFlow checkpoint from {tf_path}")
+    # Load weights from TF model
+    init_vars = tf.train.list_variables(tf_path)
+    names = []
+    arrays = []
+    for name, shape in init_vars:
+        logger.info(f"Loading TF weight {name} with shape {shape}")
+        array = tf.train.load_variable(tf_path, name)
+        names.append(name)
+        arrays.append(array.squeeze())
+
+    for name, array in zip(names, arrays):
+        name = name[6:]  # skip "model/"
+        name = name.split("/")
+        pointer = model
+        for m_name in name:
+            if re.fullmatch(r"[A-Za-z]+\d+", m_name):
+                scope_names = re.split(r"(\d+)", m_name)
+            else:
+                scope_names = [m_name]
+            if scope_names[0] == "w" or scope_names[0] == "g":
+                pointer = getattr(pointer, "weight")
+            elif scope_names[0] == "b":
+                pointer = getattr(pointer, "bias")
+            elif scope_names[0] == "wpe" or scope_names[0] == "wte":
+                pointer = getattr(pointer, scope_names[0])
+                pointer = getattr(pointer, "weight")
+            else:
+                pointer = getattr(pointer, scope_names[0])
+            if len(scope_names) >= 2:
+                num = int(scope_names[1])
+                pointer = pointer[num]
+        try:
+            assert (
+                pointer.shape == array.shape
+            ), f"Pointer shape {pointer.shape} and array shape {array.shape} mismatched"
+        except AssertionError as e:
+            e.args += (pointer.shape, array.shape)
+            raise
+        logger.info(f"Initialize PyTorch weight {name}")
+        pointer.data = torch.from_numpy(array)
+    return model
+
+
+class GPT2Attention(nn.Module):
+    def __init__(self, config, is_cross_attention=False):
+        super().__init__()
+
+        max_positions = config.max_position_embeddings
+        self.register_buffer(
+            "bias",
+            torch.tril(torch.ones((max_positions, max_positions), dtype=torch.uint8)).view(
+                1, 1, max_positions, max_positions
+            ),
+        )
+        self.register_buffer("masked_bias", torch.tensor(-1e4))
+
+        self.embed_dim = config.hidden_size
+        self.num_heads = config.num_attention_heads
+        self.head_dim = self.embed_dim // self.num_heads
+        self.split_size = self.embed_dim
+        if self.head_dim * self.num_heads != self.embed_dim:
+            raise ValueError(
+                f"`embed_dim` must be divisible by num_heads (got `embed_dim`: {self.embed_dim} and `num_heads`: {self.num_heads})."
+            )
+
+        self.scale_attn_weights = config.scale_attn_weights
+        self.is_cross_attention = is_cross_attention
+
+        if self.is_cross_attention:
+            self.c_attn = Conv1D(2 * self.embed_dim, self.embed_dim)
+            self.q_attn = Conv1D(self.embed_dim, self.embed_dim)
+        else:
+            self.c_attn = Conv1D(3 * self.embed_dim, self.embed_dim)
+        self.c_proj = Conv1D(self.embed_dim, self.embed_dim)
+
+        self.attn_dropout = nn.Dropout(config.attn_pdrop)
+        self.resid_dropout = nn.Dropout(config.resid_pdrop)
+
+        self.pruned_heads = set()
+        self.position_embedding_type = config.position_embedding_type
+
+    def prune_heads(self, heads):
+        if len(heads) == 0:
+            return
+        heads, index = find_pruneable_heads_and_indices(heads, self.num_heads, self.head_dim, self.pruned_heads)
+        index_attn = torch.cat([index, index + self.split_size, index + (2 * self.split_size)])
+
+        # Prune conv1d layers
+        self.c_attn = prune_conv1d_layer(self.c_attn, index_attn, dim=1)
+        self.c_proj = prune_conv1d_layer(self.c_proj, index, dim=0)
+
+        # Update hyper params
+        self.split_size = (self.split_size // self.num_heads) * (self.num_heads - len(heads))
+        self.num_heads = self.num_heads - len(heads)
+        self.pruned_heads = self.pruned_heads.union(heads)
+
+    def _attn(self, query, key, value, attention_mask=None, head_mask=None):
+        
+        # [b, np, sq, sk]
+        output_size = (query.size(1),
+                       query.size(2),
+                       query.size(0),
+                       key.size(0))
+        # preallocting result tensor: [b * np, sq, sk]
+        if alibi is None:
+            matmul_result = torch.empty(
+                output_size[0]*output_size[1],
+                output_size[2],
+                output_size[3],
+                dtype=query_layer.dtype,
+                device=torch.cuda.current_device())
+        else:
+            matmul_result = alibi[:output_size[0]*output_size[1], :, :output_size[3]]
+            
+        # [sq, b, np, hn] -> [sq, b * np, hn]
+        query = query.view(output_size[2],
+                                       output_size[0] * output_size[1], -1)
+        # [sk, b, np, hn] -> [sk, b * np, hn]
+        key = key.view(output_size[3],
+                                   output_size[0] * output_size[1], -1)
+        # Raw attention scores. [b * np, sq, sk]
+        attn_weights = torch.baddbmm(
+            matmul_result,
+            query_layer.transpose(0, 1),   # [b * np, sq, hn]
+            key_layer.transpose(0, 1).transpose(-1, -2),  # [b * np, hn, sk]
+            beta=0.0 if alibi is None else 1.0, alpha=(1.0/self.norm_factor))
+        
+        #attn_weights = torch.matmul(query, key.transpose(-1, -2))
+
+        # change view to [b, np, sq, sk]
+        attn_weights = attn_weights.view(*output_size)
+
+        # do we need this scaling. does the alpha do the scaling as above?
+        if self.scale_attn_weights:
+            attn_weights = attn_weights / (float(value.size(-1)) ** 0.5)
+
+        if not self.is_cross_attention:
+            # if only "normal" attention layer implements causal mask
+            query_length, key_length = query.size(-2), key.size(-2)
+            causal_mask = self.bias[:, :, key_length - query_length : key_length, :key_length].bool()
+            attn_weights = torch.where(causal_mask, attn_weights, self.masked_bias.to(attn_weights.dtype))
+
+        if attention_mask is not None:
+            # Apply the attention mask
+            attn_weights = attn_weights + attention_mask
+
+        attn_weights = nn.Softmax(dim=-1)(attn_weights)
+        attn_weights = self.attn_dropout(attn_weights)
+
+        # Mask heads if we want to
+        if head_mask is not None:
+            attn_weights = attn_weights * head_mask
+
+        attn_output = torch.matmul(attn_weights, value)
+
+        return attn_output, attn_weights
+
+    def _split_heads(self, tensor, num_heads, attn_head_size):
+        """
+        Splits hidden_size dim into attn_head_size and num_heads
+        """
+        new_shape = tensor.size()[:-1] + (num_heads, attn_head_size)
+        tensor = tensor.view(*new_shape)
+        return tensor.permute(0, 2, 1, 3)  # (batch, head, seq_length, head_features)
+
+    def _merge_heads(self, tensor, num_heads, attn_head_size):
+        """
+        Merges attn_head_size dim and num_attn_heads dim into hidden_size
+        """
+        tensor = tensor.permute(0, 2, 1, 3).contiguous()
+        new_shape = tensor.size()[:-2] + (num_heads * attn_head_size,)
+        return tensor.view(new_shape)
+
+    def forward(
+        self,
+        hidden_states,
+        layer_past=None,
+        attention_mask=None,
+        head_mask=None,
+        encoder_hidden_states=None,
+        encoder_attention_mask=None,
+        alibi=None,
+        use_cache=False,
+        output_attentions=False,
+        
+    ):
+        if encoder_hidden_states is not None:
+            if not hasattr(self, "q_attn"):
+                raise ValueError(
+                    "If class is used as cross attention, the weights `q_attn` have to be defined. "
+                    "Please make sure to instantiate class with `GPT2Attention(..., is_cross_attention=True)`."
+                )
+
+            query = self.q_attn(hidden_states)
+            key, value = self.c_attn(encoder_hidden_states).split(self.split_size, dim=2)
+            attention_mask = encoder_attention_mask
+        else:
+            query, key, value = self.c_attn(hidden_states).split(self.split_size, dim=2)
+
+        query = self._split_heads(query, self.num_heads, self.head_dim)
+        key = self._split_heads(key, self.num_heads, self.head_dim)
+        value = self._split_heads(value, self.num_heads, self.head_dim)
+
+        if layer_past is not None:
+            past_key, past_value = layer_past
+            key = torch.cat((past_key, key), dim=-2)
+            value = torch.cat((past_value, value), dim=-2)
+
+        if use_cache is True:
+            present = (key, value)
+        else:
+            present = None
+
+        attn_output, attn_weights = self._attn(query, key, value, attention_mask, head_mask)
+
+        attn_output = self._merge_heads(attn_output, self.num_heads, self.head_dim)
+        attn_output = self.c_proj(attn_output)
+        attn_output = self.resid_dropout(attn_output)
+
+        outputs = (attn_output, present)
+        if output_attentions:
+            outputs += (attn_weights,)
+
+        return outputs  # a, present, (attentions)
+
+
+class GPT2MLP(nn.Module):
+    def __init__(self, intermediate_size, config):
+        super().__init__()
+        embed_dim = config.hidden_size
+        self.c_fc = Conv1D(intermediate_size, embed_dim)
+        self.c_proj = Conv1D(embed_dim, intermediate_size)
+        self.act = ACT2FN[config.activation_function]
+        self.dropout = nn.Dropout(config.resid_pdrop)
+
+    def forward(self, hidden_states):
+        hidden_states = self.c_fc(hidden_states)
+        hidden_states = self.act(hidden_states)
+        hidden_states = self.c_proj(hidden_states)
+        hidden_states = self.dropout(hidden_states)
+        return hidden_states
+
+
+class GPT2Block(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        hidden_size = config.hidden_size
+        inner_dim = config.n_inner if config.n_inner is not None else 4 * hidden_size
+
+        self.ln_1 = nn.LayerNorm(hidden_size, eps=config.layer_norm_epsilon)
+        self.attn = GPT2Attention(config)
+        self.ln_2 = nn.LayerNorm(hidden_size, eps=config.layer_norm_epsilon)
+
+        if config.add_cross_attention:
+            self.crossattention = GPT2Attention(config, is_cross_attention=True)
+            self.ln_cross_attn = nn.LayerNorm(hidden_size, eps=config.layer_norm_epsilon)
+
+        self.mlp = GPT2MLP(inner_dim, config)
+
+    def forward(
+        self,
+        hidden_states,
+        layer_past=None,
+        attention_mask=None,
+        head_mask=None,
+        encoder_hidden_states=None,
+        encoder_attention_mask=None,
+        alibi=None,
+        use_cache=False,
+        output_attentions=False,
+    ):
+        residual = hidden_states
+        hidden_states = self.ln_1(hidden_states)
+        attn_outputs = self.attn(
+            hidden_states,
+            layer_past=layer_past,
+            attention_mask=attention_mask,
+            head_mask=head_mask,
+            alibi=alibi,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+        )
+        attn_output = attn_outputs[0]  # output_attn: a, present, (attentions)
+        outputs = attn_outputs[1:]
+        # residual connection
+        hidden_states = attn_output + residual
+
+        if encoder_hidden_states is not None:
+            # add one self-attention block for cross-attention
+            if not hasattr(self, "crossattention"):
+                raise ValueError(
+                    f"If `encoder_hidden_states` are passed, {self} has to be instantiated with "
+                    "cross-attention layers by setting `config.add_cross_attention=True`"
+                )
+            residual = hidden_states
+            hidden_states = self.ln_cross_attn(hidden_states)
+            cross_attn_outputs = self.crossattention(
+                hidden_states,
+                attention_mask=attention_mask,
+                head_mask=head_mask,
+                encoder_hidden_states=encoder_hidden_states,
+                encoder_attention_mask=encoder_attention_mask,
+                alibi=alibi,
+                output_attentions=output_attentions,
+            )
+            attn_output = cross_attn_outputs[0]
+            # residual connection
+            hidden_states = residual + attn_output
+            outputs = outputs + cross_attn_outputs[2:]  # add cross attentions if we output attention weights
+
+        residual = hidden_states
+        hidden_states = self.ln_2(hidden_states)
+        feed_forward_hidden_states = self.mlp(hidden_states)
+        # residual connection
+        hidden_states = residual + feed_forward_hidden_states
+
+        if use_cache:
+            outputs = (hidden_states,) + outputs
+        else:
+            outputs = (hidden_states,) + outputs[1:]
+
+        return outputs  # hidden_states, present, (attentions, cross_attentions)
+
+
+class GPT2PreTrainedModel(PreTrainedModel):
+    """
+    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
+    models.
+    """
+
+    config_class = GPT2Config
+    load_tf_weights = load_tf_weights_in_gpt2
+    base_model_prefix = "transformer"
+    is_parallelizable = True
+    supports_gradient_checkpointing = True
+
+    def __init__(self, *inputs, **kwargs):
+        super().__init__(*inputs, **kwargs)
+
+            
+    def _init_weights(self, module):
+        """Initialize the weights."""
+        if isinstance(module, (nn.Linear, Conv1D)):
+            # Slightly different from the TF version which uses truncated_normal for initialization
+            # cf https://github.com/pytorch/pytorch/pull/5617
+            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
+            if module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, nn.Embedding):
+            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
+            if module.padding_idx is not None:
+                module.weight.data[module.padding_idx].zero_()
+        elif isinstance(module, nn.LayerNorm):
+            module.bias.data.zero_()
+            module.weight.data.fill_(1.0)
+
+    def _set_gradient_checkpointing(self, module, value=False):
+        if isinstance(module, GPT2Model):
+            module.gradient_checkpointing = value
+
+
+
+@dataclass
+class GPT2DoubleHeadsModelOutput(ModelOutput):
+    """
+    Base class for outputs of models predicting if two sentences are consecutive or not.
+
+    Args:
+        loss (:obj:`torch.FloatTensor` of shape :obj:`(1,)`, `optional`, returned when ``labels`` is provided):
+            Language modeling loss.
+        mc_loss (:obj:`torch.FloatTensor` of shape :obj:`(1,)`, `optional`, returned when :obj:`mc_labels` is provided):
+            Multiple choice classification loss.
+        logits (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, num_choices, sequence_length, config.vocab_size)`):
+            Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
+        mc_logits (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, num_choices)`):
+            Prediction scores of the multiple choice classification head (scores for each choice before SoftMax).
+        past_key_values (:obj:`Tuple[Tuple[torch.Tensor]]`, `optional`, returned when ``use_cache=True`` is passed or when ``config.use_cache=True``):
+            Tuple of length :obj:`config.n_layers`, containing tuples of tensors of shape :obj:`(batch_size, num_heads,
+            sequence_length, embed_size_per_head)`).
+
+            Contains pre-computed hidden-states (key and values in the attention blocks) that can be used (see
+            :obj:`past_key_values` input) to speed up sequential decoding.
+        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
+            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape :obj:`(batch_size, num_heads,
+            sequence_length, sequence_length)`.
+
+            GPT2Attentions weights after the attention softmax, used to compute the weighted average in the
+            self-attention heads.
+    """
+
+    loss: Optional[torch.FloatTensor] = None
+    mc_loss: Optional[torch.FloatTensor] = None
+    logits: torch.FloatTensor = None
+    mc_logits: torch.FloatTensor = None
+    past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None
+    hidden_states: Optional[Tuple[torch.FloatTensor]] = None
+    attentions: Optional[Tuple[torch.FloatTensor]] = None
+
+
+GPT2_START_DOCSTRING = r"""
+
+    This model inherits from :class:`~transformers.PreTrainedModel`. Check the superclass documentation for the generic
+    methods the library implements for all its model (such as downloading or saving, resizing the input embeddings,
+    pruning heads etc.)
+
+    This model is also a PyTorch `torch.nn.Module <https://pytorch.org/docs/stable/nn.html#torch.nn.Module>`__
+    subclass. Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to
+    general usage and behavior.
+
+    Parameters:
+        config (:class:`~transformers.GPT2Config`): Model configuration class with all the parameters of the model.
+            Initializing with a config file does not load the weights associated with the model, only the
+            configuration. Check out the :meth:`~transformers.PreTrainedModel.from_pretrained` method to load the model
+            weights.
+"""
+
+GPT2_INPUTS_DOCSTRING = r"""
+    Args:
+        input_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size, input_ids_length)`):
+            :obj:`input_ids_length` = ``sequence_length`` if :obj:`past_key_values` is ``None`` else
+            ``past_key_values[0][0].shape[-2]`` (``sequence_length`` of input past key value states). Indices of input
+            sequence tokens in the vocabulary.
+
+            If :obj:`past_key_values` is used, only ``input_ids`` that do not have their past calculated should be
+            passed as ``input_ids``.
+
+            Indices can be obtained using :class:`~transformers.GPT2Tokenizer`. See
+            :meth:`transformers.PreTrainedTokenizer.encode` and :meth:`transformers.PreTrainedTokenizer.__call__` for
+            details.
+
+            `What are input IDs? <../glossary.html#input-ids>`__
+        past_key_values (:obj:`Tuple[Tuple[torch.Tensor]]` of length :obj:`config.n_layers`):
+            Contains precomputed hidden-states (key and values in the attention blocks) as computed by the model (see
+            :obj:`past_key_values` output below). Can be used to speed up sequential decoding. The ``input_ids`` which
+            have their past given to this model should not be passed as ``input_ids`` as they have already been
+            computed.
+        attention_mask (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):
+            Mask to avoid performing attention on padding token indices. Mask values selected in ``[0, 1]``:
+
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+
+            `What are attention masks? <../glossary.html#attention-mask>`__
+        token_type_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size, input_ids_length)`, `optional`):
+            Segment token indices to indicate first and second portions of the inputs. Indices are selected in ``[0,
+            1]``:
+
+            - 0 corresponds to a `sentence A` token,
+            - 1 corresponds to a `sentence B` token.
+
+            `What are token type IDs? <../glossary.html#token-type-ids>`_
+        position_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):
+            Indices of positions of each input sequence tokens in the position embeddings. Selected in the range ``[0,
+            config.max_position_embeddings - 1]``.
+
+            `What are position IDs? <../glossary.html#position-ids>`_
+        head_mask (:obj:`torch.FloatTensor` of shape :obj:`(num_heads,)` or :obj:`(num_layers, num_heads)`, `optional`):
+            Mask to nullify selected heads of the self-attention modules. Mask values selected in ``[0, 1]``:
+
+            - 1 indicates the head is **not masked**,
+            - 0 indicates the head is **masked**.
+
+        inputs_embeds (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`, `optional`):
+            Optionally, instead of passing :obj:`input_ids` you can choose to directly pass an embedded representation.
+            This is useful if you want more control over how to convert :obj:`input_ids` indices into associated
+            vectors than the model's internal embedding lookup matrix.
+
+            If :obj:`past_key_values` is used, optionally only the last :obj:`inputs_embeds` have to be input (see
+            :obj:`past_key_values`).
+        use_cache (:obj:`bool`, `optional`):
+            If set to :obj:`True`, :obj:`past_key_values` key value states are returned and can be used to speed up
+            decoding (see :obj:`past_key_values`).
+        output_attentions (:obj:`bool`, `optional`):
+            Whether or not to return the attentions tensors of all attention layers. See ``attentions`` under returned
+            tensors for more detail.
+        output_hidden_states (:obj:`bool`, `optional`):
+            Whether or not to return the hidden states of all layers. See ``hidden_states`` under returned tensors for
+            more detail.
+        return_dict (:obj:`bool`, `optional`):
+            Whether or not to return a :class:`~transformers.file_utils.ModelOutput` instead of a plain tuple.
+"""
+PARALLELIZE_DOCSTRING = r"""
+    This is an experimental feature and is a subject to change at a moment's notice.
+
+    Uses a device map to distribute attention modules of the model across several devices. If no device map is given,
+    it will evenly distribute blocks across all devices.
+
+    Args:
+        device_map (:obj:`Dict[int, list]`, optional, defaults to None):
+            A dictionary that maps attention modules to devices. Note that the embedding module and LMHead are always
+            automatically mapped to the first device (for esoteric reasons). That means that the first device should
+            have fewer attention modules mapped to it than other devices. For reference, the gpt2 models have the
+            following number of attention modules:
+
+                - gpt2: 12
+                - gpt2-medium: 24
+                - gpt2-large: 36
+                - gpt2-xl: 48
+
+    Example::
+
+            # Here is an example of a device map on a machine with 4 GPUs using gpt2-xl, which has a total of 48 attention modules:
+            model = GPT2LMHeadModel.from_pretrained('gpt2-xl')
+            device_map = {0: [0, 1, 2, 3, 4, 5, 6, 7, 8],
+
+                          1: [9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21],
+                          2: [22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34],
+                          3: [35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47]}
+            model.parallelize(device_map)
+"""
+DEPARALLELIZE_DOCSTRING = r"""
+    Moves the model to cpu from a model parallel state.
+
+    Example::
+
+        # On a 4 GPU machine with gpt2-large:
+        model = GPT2LMHeadModel.from_pretrained('gpt2-large')
+        device_map = {0: [0, 1, 2, 3, 4, 5, 6, 7],
+
+                    1: [8, 9, 10, 11, 12, 13, 14, 15],
+                    2: [16, 17, 18, 19, 20, 21, 22, 23],
+                    3: [24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35]}
+        model.parallelize(device_map) # Splits the model across several devices
+        model.deparallelize() # Put the model back on cpu and cleans memory by calling torch.cuda.empty_cache()
+"""
+
+
+@add_start_docstrings(
+    "The bare GPT2 Model transformer outputting raw hidden-states without any specific head on top.",
+    GPT2_START_DOCSTRING,
+)
+class GPT2Model(GPT2PreTrainedModel):
+    _keys_to_ignore_on_load_missing = ["attn.masked_bias"]
+
+    def __init__(self, config):
+        super().__init__(config)
+
+        self.embed_dim = config.hidden_size
+
+        self.wte = nn.Embedding(config.vocab_size, self.embed_dim)
+        self.wpe = nn.Embedding(config.max_position_embeddings, self.embed_dim)
+
+        self.drop = nn.Dropout(config.embd_pdrop)
+        self.h = nn.ModuleList([GPT2Block(config) for _ in range(config.num_hidden_layers)])
+        self.ln_f = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_epsilon)
+
+        self.init_weights()
+
+        # Model parallel
+        self.model_parallel = False
+        self.device_map = None
+        self.gradient_checkpointing = False
+        config = kwargs.get('config',inputs[0])
+        if args.position_embedding_type == PositionEmbeddingType_alibi:
+            self.alibi = self._build_alibi_tensor(args.seq_length, args.num_attention_heads, args.micro_batch_size).to(torch.cuda.current_device())
+            if args.params_dtype == torch.float16:
+                self.alibi = self.alibi.to(torch.float16)
+            elif args.params_dtype == torch.bfloat16:
+                self.alibi = self.alibi.to(torch.bfloat16)
+        else:
+            self.alibi = None
+
+    @staticmethod
+    def _build_alibi_tensor(max_seq_len, num_attention_heads, batch_size):
+        # Based on https://github.com/ofirpress/attention_with_linear_biases/blob/a35aaca144e0eb6b789dfcb46784c4b8e31b7983/fairseq/models/transformer.py#L742
+        """Returns tensor shaped (batch_size * num_attention_heads, 1, max_seq_len)"""
+        def get_slopes(n):
+            def get_slopes_power_of_2(n):
+                start = (2 ** (-2 ** -(math.log2(n) - 3)))
+                ratio = start
+                return [start * ratio ** i for i in range(n)]
+
+            if math.log2(n).is_integer():
+                return get_slopes_power_of_2(n)
+            else:
+                closest_power_of_2 = 2 ** math.floor(math.log2(n))
+                return get_slopes_power_of_2(closest_power_of_2) + get_slopes(2 * closest_power_of_2)[0::2][
+                                                                   :n - closest_power_of_2]
+        slopes = torch.Tensor(get_slopes(num_attention_heads))
+        alibi = slopes.unsqueeze(1).unsqueeze(1) * torch.arange(max_seq_len).unsqueeze(0).unsqueeze(0).expand(num_attention_heads, -1, -1)
+        alibi = alibi.repeat(batch_size, 1, 1)
+        return alibi
+
+
+    @add_start_docstrings(PARALLELIZE_DOCSTRING)
+    def parallelize(self, device_map=None):
+        # Check validity of device_map
+        self.device_map = (
+            get_device_map(len(self.h), range(torch.cuda.device_count())) if device_map is None else device_map
+        )
+        assert_device_map(self.device_map, len(self.h))
+        self.model_parallel = True
+        self.first_device = "cpu" if "cpu" in self.device_map.keys() else "cuda:" + str(min(self.device_map.keys()))
+        self.last_device = "cuda:" + str(max(self.device_map.keys()))
+        self.wte = self.wte.to(self.first_device)
+        self.wpe = self.wpe.to(self.first_device)
+        # Load onto devices
+        for k, v in self.device_map.items():
+            for block in v:
+                cuda_device = "cuda:" + str(k)
+                self.h[block] = self.h[block].to(cuda_device)
+        # ln_f to last
+        self.ln_f = self.ln_f.to(self.last_device)
+
+    @add_start_docstrings(DEPARALLELIZE_DOCSTRING)
+    def deparallelize(self):
+        self.model_parallel = False
+        self.device_map = None
+        self.first_device = "cpu"
+        self.last_device = "cpu"
+        self.wte = self.wte.to("cpu")
+        self.wpe = self.wpe.to("cpu")
+        for index in range(len(self.h)):
+            self.h[index] = self.h[index].to("cpu")
+        self.ln_f = self.ln_f.to("cpu")
+        torch.cuda.empty_cache()
+
+    def get_input_embeddings(self):
+        return self.wte
+
+    def set_input_embeddings(self, new_embeddings):
+        self.wte = new_embeddings
+
+    def _prune_heads(self, heads_to_prune):
+        """
+        Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer}
+        """
+        for layer, heads in heads_to_prune.items():
+            self.h[layer].attn.prune_heads(heads)
+
+    @add_start_docstrings_to_model_forward(GPT2_INPUTS_DOCSTRING)
+    @add_code_sample_docstrings(
+        tokenizer_class=_TOKENIZER_FOR_DOC,
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=BaseModelOutputWithPastAndCrossAttentions,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    def forward(
+        self,
+        input_ids=None,
+        past_key_values=None,
+        attention_mask=None,
+        token_type_ids=None,
+        position_ids=None,
+        head_mask=None,
+        inputs_embeds=None,
+        encoder_hidden_states=None,
+        encoder_attention_mask=None,
+        use_cache=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+        prefix_lm_token_id = None
+    ):
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        use_cache = use_cache if use_cache is not None else self.config.use_cache
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
+        elif input_ids is not None:
+            input_shape = input_ids.size()
+            input_ids = input_ids.view(-1, input_shape[-1])
+            batch_size = input_ids.shape[0]
+        elif inputs_embeds is not None:
+            input_shape = inputs_embeds.size()[:-1]
+            batch_size = inputs_embeds.shape[0]
+        else:
+            raise ValueError("You have to specify either input_ids or inputs_embeds")
+
+        device = input_ids.device if input_ids is not None else inputs_embeds.device
+
+        if token_type_ids is not None:
+            token_type_ids = token_type_ids.view(-1, input_shape[-1])
+        if position_ids is not None:
+            position_ids = position_ids.view(-1, input_shape[-1])
+
+        if past_key_values is None:
+            past_length = 0
+            past_key_values = tuple([None] * len(self.h))
+        else:
+            past_length = past_key_values[0][0].size(-2)
+        if position_ids is None:
+            position_ids = torch.arange(past_length, input_shape[-1] + past_length, dtype=torch.long, device=device)
+            position_ids = position_ids.unsqueeze(0).view(-1, input_shape[-1])
+
+        # GPT2Attention mask.
+        if attention_mask is not None:
+            if batch_size <= 0:
+                raise ValueError("batch_size has to be defined and > 0")
+            attention_mask = attention_mask.view(batch_size, -1)
+            # do prefix_lm masking if we have input_ids. We find the prefix_lm_toke_id token as the prefix_lm boundry.
+            if prefix_lm_token_id is not None and input_ids is not None:
+                for attention_mask_row, input_ids_row in zip(attention_mask, input_ids): # do this in the bs dimension
+                    attention_mask_row[: (input_ids_row == prefix_lm_token_id).nonzero(as_tuple=True)[0], :] = 1.0 # is this right?
+                
+            # We create a 3D attention mask from a 2D tensor mask.
+            # Sizes are [batch_size, 1, 1, to_seq_length]
+            # So we can broadcast to [batch_size, num_heads, from_seq_length, to_seq_length]
+            # this attention mask is more simple than the triangular masking of causal attention
+            # used in OpenAI GPT, we just need to prepare the broadcast dimension here.
+            attention_mask = attention_mask[:, None, None, :]
+
+            # Since attention_mask is 1.0 for positions we want to attend and 0.0 for
+            # masked positions, this operation will create a tensor which is 0.0 for
+            # positions we want to attend and -10000.0 for masked positions.
+            # Since we are adding it to the raw scores before the softmax, this is
+            # effectively the same as removing these entirely.
+            attention_mask = attention_mask.to(dtype=self.dtype)  # fp16 compatibility
+            attention_mask = (1.0 - attention_mask) * -10000.0
+
+        # If a 2D ou 3D attention mask is provided for the cross-attention
+        # we need to make broadcastable to [batch_size, num_heads, seq_length, seq_length]
+        if self.config.add_cross_attention and encoder_hidden_states is not None:
+            encoder_batch_size, encoder_sequence_length, _ = encoder_hidden_states.size()
+            encoder_hidden_shape = (encoder_batch_size, encoder_sequence_length)
+            if encoder_attention_mask is None:
+                encoder_attention_mask = torch.ones(encoder_hidden_shape, device=device)
+            encoder_attention_mask = self.invert_attention_mask(encoder_attention_mask)
+        else:
+            encoder_attention_mask = None
+
+        # Prepare head mask if needed
+        # 1.0 in head_mask indicate we keep the head
+        # attention_probs has shape bsz x n_heads x N x N
+        # head_mask has shape n_layer x batch x n_heads x N x N
+        head_mask = self.get_head_mask(head_mask, self.config.n_layer)
+
+        if inputs_embeds is None:
+            inputs_embeds = self.wte(input_ids)
+        position_embeds = self.wpe(position_ids)
+        hidden_states = inputs_embeds + position_embeds
+
+        if token_type_ids is not None:
+            token_type_embeds = self.wte(token_type_ids)
+            hidden_states = hidden_states + token_type_embeds
+
+        hidden_states = self.drop(hidden_states)
+
+        output_shape = input_shape + (hidden_states.size(-1),)
+
+        presents = () if use_cache else None
+        all_self_attentions = () if output_attentions else None
+        all_cross_attentions = () if output_attentions and self.config.add_cross_attention else None
+        all_hidden_states = () if output_hidden_states else None
+        for i, (block, layer_past) in enumerate(zip(self.h, past_key_values)):
+
+            # Model parallel
+            if self.model_parallel:
+                torch.cuda.set_device(hidden_states.device)
+                # Ensure layer_past is on same device as hidden_states (might not be correct)
+                if layer_past is not None:
+                    layer_past = tuple(past_state.to(hidden_states.device) for past_state in layer_past)
+                # Ensure that attention_mask is always on the same device as hidden_states
+                if attention_mask is not None:
+                    attention_mask = attention_mask.to(hidden_states.device)
+                if isinstance(head_mask, torch.Tensor):
+                    head_mask = head_mask.to(hidden_states.device)
+            if output_hidden_states:
+                all_hidden_states = all_hidden_states + (hidden_states,)
+
+            if self.gradient_checkpointing and self.training:
+
+                if use_cache:
+                    logger.warning(
+                        "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
+                    )
+                    use_cache = False
+
+                def create_custom_forward(module):
+                    def custom_forward(*inputs):
+                        # None for past_key_value
+                        return module(*inputs, use_cache, output_attentions)
+
+                    return custom_forward
+
+                outputs = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(block),
+                    hidden_states,
+                    None,
+                    attention_mask,
+                    head_mask[i],
+                    encoder_hidden_states,
+                    encoder_attention_mask,
+                    self.alibi
+                )
+            else:
+                outputs = block(
+                    hidden_states,
+                    layer_past=layer_past,
+                    attention_mask=attention_mask,
+                    head_mask=head_mask[i],
+                    encoder_hidden_states=encoder_hidden_states,
+                    encoder_attention_mask=encoder_attention_mask,
+                    use_cache=use_cache,
+                    output_attentions=output_attentions,
+                    alibi=self.alibi
+                )
+
+            hidden_states = outputs[0]
+            if use_cache is True:
+                presents = presents + (outputs[1],)
+
+            if output_attentions:
+                all_self_attentions = all_self_attentions + (outputs[2 if use_cache else 1],)
+                if self.config.add_cross_attention:
+                    all_cross_attentions = all_cross_attentions + (outputs[3 if use_cache else 2],)
+
+            # Model Parallel: If it's the last layer for that device, put things on the next device
+            if self.model_parallel:
+                for k, v in self.device_map.items():
+                    if i == v[-1] and "cuda:" + str(k) != self.last_device:
+                        hidden_states = hidden_states.to("cuda:" + str(k + 1))
+
+        hidden_states = self.ln_f(hidden_states)
+
+        hidden_states = hidden_states.view(*output_shape)
+        # Add last hidden state
+        if output_hidden_states:
+            all_hidden_states = all_hidden_states + (hidden_states,)
+
+        if not return_dict:
+            return tuple(
+                v
+                for v in [hidden_states, presents, all_hidden_states, all_self_attentions, all_cross_attentions]
+                if v is not None
+            )
+
+        return BaseModelOutputWithPastAndCrossAttentions(
+            last_hidden_state=hidden_states,
+            past_key_values=presents,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attentions,
+            cross_attentions=all_cross_attentions,
+        )
+
+
+@add_start_docstrings(
+    """
+    The GPT2 Model transformer with a language modeling head on top (linear layer with weights tied to the input
+    embeddings).
+    """,
+    GPT2_START_DOCSTRING,
+)
+class GPT2LMHeadModel(GPT2PreTrainedModel):
+    _keys_to_ignore_on_load_missing = [r"attn.masked_bias", r"attn.bias", r"lm_head.weight"]
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.transformer = GPT2Model(config)
+        self.lm_head = nn.Linear(config.n_embd, config.vocab_size, bias=False)
+
+        self.init_weights()
+
+        # Model parallel
+        self.model_parallel = False
+        self.device_map = None
+
+    @add_start_docstrings(PARALLELIZE_DOCSTRING)
+    def parallelize(self, device_map=None):
+        self.device_map = (
+            get_device_map(len(self.transformer.h), range(torch.cuda.device_count()))
+            if device_map is None
+            else device_map
+        )
+        assert_device_map(self.device_map, len(self.transformer.h))
+        self.transformer.parallelize(self.device_map)
+        self.lm_head = self.lm_head.to(self.transformer.first_device)
+        self.model_parallel = True
+
+    @add_start_docstrings(DEPARALLELIZE_DOCSTRING)
+    def deparallelize(self):
+        self.transformer.deparallelize()
+        self.transformer = self.transformer.to("cpu")
+        self.lm_head = self.lm_head.to("cpu")
+        self.model_parallel = False
+        torch.cuda.empty_cache()
+
+    def get_output_embeddings(self):
+        return self.lm_head
+
+    def set_output_embeddings(self, new_embeddings):
+        self.lm_head = new_embeddings
+
+    def prepare_inputs_for_generation(self, input_ids, past=None, **kwargs):
+        token_type_ids = kwargs.get("token_type_ids", None)
+        # only last token for inputs_ids if past is defined in kwargs
+        if past:
+            input_ids = input_ids[:, -1].unsqueeze(-1)
+            if token_type_ids is not None:
+                token_type_ids = token_type_ids[:, -1].unsqueeze(-1)
+
+        attention_mask = kwargs.get("attention_mask", None)
+        position_ids = kwargs.get("position_ids", None)
+
+        if attention_mask is not None and position_ids is None:
+            # create position_ids on the fly for batch generation
+            position_ids = attention_mask.long().cumsum(-1) - 1
+            position_ids.masked_fill_(attention_mask == 0, 1)
+            if past:
+                position_ids = position_ids[:, -1].unsqueeze(-1)
+        else:
+            position_ids = None
+        return {
+            "input_ids": input_ids,
+            "past_key_values": past,
+            "use_cache": kwargs.get("use_cache"),
+            "position_ids": position_ids,
+            "attention_mask": attention_mask,
+            "token_type_ids": token_type_ids,
+        }
+
+    @add_start_docstrings_to_model_forward(GPT2_INPUTS_DOCSTRING)
+    @add_code_sample_docstrings(
+        tokenizer_class=_TOKENIZER_FOR_DOC,
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=CausalLMOutputWithCrossAttentions,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    def forward(
+        self,
+        input_ids=None,
+        past_key_values=None,
+        attention_mask=None,
+        token_type_ids=None,
+        position_ids=None,
+        head_mask=None,
+        inputs_embeds=None,
+        encoder_hidden_states=None,
+        encoder_attention_mask=None,
+        labels=None,
+        use_cache=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+    ):
+        r"""
+        labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):
+            Labels for language modeling. Note that the labels **are shifted** inside the model, i.e. you can set
+            ``labels = input_ids`` Indices are selected in ``[-100, 0, ..., config.vocab_size]`` All labels set to
+            ``-100`` are ignored (masked), the loss is only computed for labels in ``[0, ..., config.vocab_size]``
+        """
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        transformer_outputs = self.transformer(
+            input_ids,
+            past_key_values=past_key_values,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=encoder_attention_mask,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+        hidden_states = transformer_outputs[0]
+
+        # Set device for model parallelism
+        if self.model_parallel:
+            torch.cuda.set_device(self.transformer.first_device)
+            hidden_states = hidden_states.to(self.lm_head.weight.device)
+
+        lm_logits = self.lm_head(hidden_states)
+
+        loss = None
+        if labels is not None:
+            # Shift so that tokens < n predict n
+            shift_logits = lm_logits[..., :-1, :].contiguous()
+            shift_labels = labels[..., 1:].contiguous()
+            # Flatten the tokens
+            loss_fct = CrossEntropyLoss()
+            loss = loss_fct(shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1))
+
+        if not return_dict:
+            output = (lm_logits,) + transformer_outputs[1:]
+            return ((loss,) + output) if loss is not None else output
+
+        return CausalLMOutputWithCrossAttentions(
+            loss=loss,
+            logits=lm_logits,
+            past_key_values=transformer_outputs.past_key_values,
+            hidden_states=transformer_outputs.hidden_states,
+            attentions=transformer_outputs.attentions,
+            cross_attentions=transformer_outputs.cross_attentions,
+        )
+
+    @staticmethod
+    def _reorder_cache(past: Tuple[Tuple[torch.Tensor]], beam_idx: torch.Tensor) -> Tuple[Tuple[torch.Tensor]]:
+        """
+        This function is used to re-order the :obj:`past_key_values` cache if
+        :meth:`~transformers.PreTrainedModel.beam_search` or :meth:`~transformers.PreTrainedModel.beam_sample` is
+        called. This is required to match :obj:`past_key_values` with the correct beam_idx at every generation step.
+        """
+        return tuple(
+            tuple(past_state.index_select(0, beam_idx.to(past_state.device)) for past_state in layer_past)
+            for layer_past in past
+        )
+
+
+@add_start_docstrings(
+    """
+The GPT2 Model transformer with a language modeling and a multiple-choice classification head on top e.g. for
+RocStories/SWAG tasks. The two heads are two linear layers. The language modeling head has its weights tied to the
+input embeddings, the classification head takes as input the input of a specified classification token index in the
+input sequence).
+""",
+    GPT2_START_DOCSTRING,
+)
+class GPT2DoubleHeadsModel(GPT2PreTrainedModel):
+    _keys_to_ignore_on_load_missing = [r"attn.masked_bias", r"attn.bias", r"lm_head.weight"]
+
+    def __init__(self, config):
+        super().__init__(config)
+        config.num_labels = 1
+        self.transformer = GPT2Model(config)
+        self.lm_head = nn.Linear(config.n_embd, config.vocab_size, bias=False)
+        self.multiple_choice_head = SequenceSummary(config)
+
+        self.init_weights()
+
+        # Model parallel
+        self.model_parallel = False
+        self.device_map = None
+
+    @add_start_docstrings(PARALLELIZE_DOCSTRING)
+    def parallelize(self, device_map=None):
+        self.device_map = (
+            get_device_map(len(self.transformer.h), range(torch.cuda.device_count()))
+            if device_map is None
+            else device_map
+        )
+        assert_device_map(self.device_map, len(self.transformer.h))
+        self.transformer.parallelize(self.device_map)
+        self.lm_head = self.lm_head.to(self.transformer.first_device)
+        self.multiple_choice_head = self.multiple_choice_head.to(self.transformer.first_device)
+        self.model_parallel = True
+
+    @add_start_docstrings(DEPARALLELIZE_DOCSTRING)
+    def deparallelize(self):
+        self.transformer.deparallelize()
+        self.transformer = self.transformer.to("cpu")
+        self.lm_head = self.lm_head.to("cpu")
+        self.multiple_choice_head = self.multiple_choice_head.to("cpu")
+        self.model_parallel = False
+        torch.cuda.empty_cache()
+
+    def get_output_embeddings(self):
+        return self.lm_head
+
+    def set_output_embeddings(self, new_embeddings):
+        self.lm_head = new_embeddings
+
+    def prepare_inputs_for_generation(self, input_ids, past=None, **kwargs):
+        token_type_ids = kwargs.get("token_type_ids", None)
+        # only last token for inputs_ids if past is defined in kwargs
+        if past:
+            input_ids = input_ids[:, -1].unsqueeze(-1)
+            if token_type_ids is not None:
+                token_type_ids = token_type_ids[:, -1].unsqueeze(-1)
+
+        attention_mask = kwargs.get("attention_mask", None)
+        position_ids = kwargs.get("position_ids", None)
+
+        if attention_mask is not None and position_ids is None:
+            # create position_ids on the fly for batch generation
+            position_ids = attention_mask.long().cumsum(-1) - 1
+            position_ids.masked_fill_(attention_mask == 0, 1)
+            if past:
+                position_ids = position_ids[:, -1].unsqueeze(-1)
+        else:
+            position_ids = None
+
+        return {
+            "input_ids": input_ids,
+            "past_key_values": past,
+            "use_cache": kwargs.get("use_cache"),
+            "position_ids": position_ids,
+            "attention_mask": attention_mask,
+            "token_type_ids": token_type_ids,
+        }
+
+    @add_start_docstrings_to_model_forward(GPT2_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=GPT2DoubleHeadsModelOutput, config_class=_CONFIG_FOR_DOC)
+    def forward(
+        self,
+        input_ids=None,
+        past_key_values=None,
+        attention_mask=None,
+        token_type_ids=None,
+        position_ids=None,
+        head_mask=None,
+        inputs_embeds=None,
+        mc_token_ids=None,
+        labels=None,
+        mc_labels=None,
+        use_cache=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+        **kwargs,
+    ):
+        r"""
+        mc_token_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size, num_choices)`, `optional`, default to index of the last token of the input):
+            Index of the classification token in each input sequence. Selected in the range ``[0, input_ids.size(-1) -
+            1[``.
+        labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):
+            Labels for language modeling. Note that the labels **are shifted** inside the model, i.e. you can set
+            ``labels = input_ids`` Indices are selected in ``[-100, 0, ..., config.vocab_size - 1]`` All labels set to
+            ``-100`` are ignored (masked), the loss is only computed for labels in ``[0, ..., config.vocab_size - 1]``
+        mc_labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size)`, `optional`):
+            Labels for computing the multiple choice classification loss. Indices should be in ``[0, ...,
+            num_choices]`` where `num_choices` is the size of the second dimension of the input tensors. (see
+            `input_ids` above)
+
+        Return:
+
+        Example::
+
+            >>> import torch
+            >>> from transformers import GPT2Tokenizer, GPT2DoubleHeadsModel
+
+            >>> tokenizer = GPT2Tokenizer.from_pretrained('gpt2')
+            >>> model = GPT2DoubleHeadsModel.from_pretrained('gpt2')
+
+            >>> # Add a [CLS] to the vocabulary (we should train it also!)
+            >>> num_added_tokens = tokenizer.add_special_tokens({'cls_token': '[CLS]'})
+
+            >>> embedding_layer = model.resize_token_embeddings(len(tokenizer))  # Update the model embeddings with the new vocabulary size
+
+            >>> choices = ["Hello, my dog is cute [CLS]", "Hello, my cat is cute [CLS]"]
+            >>> encoded_choices = [tokenizer.encode(s) for s in choices]
+            >>> cls_token_location = [tokens.index(tokenizer.cls_token_id) for tokens in encoded_choices]
+
+            >>> input_ids = torch.tensor(encoded_choices).unsqueeze(0)  # Batch size: 1, number of choices: 2
+            >>> mc_token_ids = torch.tensor([cls_token_location])  # Batch size: 1
+
+            >>> outputs = model(input_ids, mc_token_ids=mc_token_ids)
+            >>> lm_logits = outputs.logits
+            >>> mc_logits = outputs.mc_logits
+
+        """
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        transformer_outputs = self.transformer(
+            input_ids,
+            past_key_values=past_key_values,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        hidden_states = transformer_outputs[0]
+
+        # Set device for model parallelism
+        if self.model_parallel:
+            torch.cuda.set_device(self.transformer.first_device)
+            hidden_states = hidden_states.to(self.lm_head.weight.device)
+
+        lm_logits = self.lm_head(hidden_states)
+        mc_logits = self.multiple_choice_head(hidden_states, mc_token_ids).squeeze(-1)
+
+        mc_loss = None
+        if mc_labels is not None:
+            loss_fct = CrossEntropyLoss()
+            mc_loss = loss_fct(mc_logits.view(-1, mc_logits.size(-1)), mc_labels.view(-1))
+        lm_loss = None
+        if labels is not None:
+            shift_logits = lm_logits[..., :-1, :].contiguous()
+            shift_labels = labels[..., 1:].contiguous()
+            loss_fct = CrossEntropyLoss()
+            lm_loss = loss_fct(shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1))
+
+        if not return_dict:
+            output = (lm_logits, mc_logits) + transformer_outputs[1:]
+            if mc_loss is not None:
+                output = (mc_loss,) + output
+            return ((lm_loss,) + output) if lm_loss is not None else output
+
+        return GPT2DoubleHeadsModelOutput(
+            loss=lm_loss,
+            mc_loss=mc_loss,
+            logits=lm_logits,
+            mc_logits=mc_logits,
+            past_key_values=transformer_outputs.past_key_values,
+            hidden_states=transformer_outputs.hidden_states,
+            attentions=transformer_outputs.attentions,
+        )
+
+    @staticmethod
+    def _reorder_cache(past: Tuple[Tuple[torch.Tensor]], beam_idx: torch.Tensor) -> Tuple[Tuple[torch.Tensor]]:
+        """
+        This function is used to re-order the :obj:`past_key_values` cache if
+        :meth:`~transformers.PreTrainedModel.beam_search` or :meth:`~transformers.PreTrainedModel.beam_sample` is
+        called. This is required to match :obj:`past_key_values` with the correct beam_idx at every generation step.
+        """
+        return tuple(
+            tuple(past_state.index_select(0, beam_idx.to(past_state.device)) for past_state in layer_past)
+            for layer_past in past
+        )
+
+
+@add_start_docstrings(
+    """
+    The GPT2 Model transformer with a sequence classification head on top (linear layer).
+
+    :class:`~transformers.GPT2ForSequenceClassification` uses the last token in order to do the classification, as
+    other causal models (e.g. GPT-1) do.
+
+    Since it does classification on the last token, it requires to know the position of the last token. If a
+    :obj:`pad_token_id` is defined in the configuration, it finds the last token that is not a padding token in each
+    row. If no :obj:`pad_token_id` is defined, it simply takes the last value in each row of the batch. Since it cannot
+    guess the padding tokens when :obj:`inputs_embeds` are passed instead of :obj:`input_ids`, it does the same (take
+    the last value in each row of the batch).
+    """,
+    GPT2_START_DOCSTRING,
+)
+class GPT2ForSequenceClassification(GPT2PreTrainedModel):
+    _keys_to_ignore_on_load_missing = [r"h\.\d+\.attn\.masked_bias", r"lm_head\.weight"]
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+        self.transformer = GPT2Model(config)
+        self.score = nn.Linear(config.n_embd, self.num_labels, bias=False)
+
+        self.init_weights()
+
+        # Model parallel
+        self.model_parallel = False
+        self.device_map = None
+
+    @add_start_docstrings_to_model_forward(GPT2_INPUTS_DOCSTRING)
+    @add_code_sample_docstrings(
+        tokenizer_class=_TOKENIZER_FOR_DOC,
+        checkpoint="microsoft/DialogRPT-updown",
+        output_type=SequenceClassifierOutputWithPast,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    def forward(
+        self,
+        input_ids=None,
+        past_key_values=None,
+        attention_mask=None,
+        token_type_ids=None,
+        position_ids=None,
+        head_mask=None,
+        inputs_embeds=None,
+        labels=None,
+        use_cache=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+    ):
+        r"""
+        labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size,)`, `optional`):
+            Labels for computing the sequence classification/regression loss. Indices should be in :obj:`[0, ...,
+            config.num_labels - 1]`. If :obj:`config.num_labels == 1` a regression loss is computed (Mean-Square loss),
+            If :obj:`config.num_labels > 1` a classification loss is computed (Cross-Entropy).
+        """
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        transformer_outputs = self.transformer(
+            input_ids,
+            past_key_values=past_key_values,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+        hidden_states = transformer_outputs[0]
+        logits = self.score(hidden_states)
+
+        if input_ids is not None:
+            batch_size, sequence_length = input_ids.shape[:2]
+        else:
+            batch_size, sequence_length = inputs_embeds.shape[:2]
+
+        assert (
+            self.config.pad_token_id is not None or batch_size == 1
+        ), "Cannot handle batch sizes > 1 if no padding token is defined."
+        if self.config.pad_token_id is None:
+            sequence_lengths = -1
+        else:
+            if input_ids is not None:
+                sequence_lengths = torch.ne(input_ids, self.config.pad_token_id).sum(-1) - 1
+            else:
+                sequence_lengths = -1
+                logger.warning(
+                    f"{self.__class__.__name__} will not detect padding tokens in `inputs_embeds`. Results may be "
+                    f"unexpected if using padding tokens in conjunction with `inputs_embeds.`"
+                )
+
+        pooled_logits = logits[range(batch_size), sequence_lengths]
+
+        loss = None
+        if labels is not None:
+            if self.num_labels == 1:
+                #  We are doing regression
+                loss_fct = MSELoss()
+                loss = loss_fct(pooled_logits.view(-1), labels.to(self.dtype).view(-1))
+            else:
+                loss_fct = CrossEntropyLoss()
+                loss = loss_fct(pooled_logits.view(-1, self.num_labels), labels.view(-1))
+
+        if not return_dict:
+            output = (pooled_logits,) + transformer_outputs[1:]
+            return ((loss,) + output) if loss is not None else output
+
+        return SequenceClassifierOutputWithPast(
+            loss=loss,
+            logits=pooled_logits,
+            past_key_values=transformer_outputs.past_key_values,
+            hidden_states=transformer_outputs.hidden_states,
+            attentions=transformer_outputs.attentions,
+        )
+
+
+@add_start_docstrings(
+    """
+    GPT2 Model with a token classification head on top (a linear layer on top of the hidden-states output) e.g. for
+    Named-Entity-Recognition (NER) tasks.
+    """,
+    GPT2_START_DOCSTRING,
+)
+class GPT2ForTokenClassification(GPT2PreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+
+        self.transformer = GPT2Model(config)
+        if hasattr(config, "classifier_dropout") and config.classifier_dropout is not None:
+            classifier_dropout = config.classifier_dropout
+        elif hasattr(config, "hidden_dropout") and config.hidden_dropout is not None:
+            classifier_dropout = config.hidden_dropout
+        else:
+            classifier_dropout = 0.1
+        self.dropout = nn.Dropout(classifier_dropout)
+        self.classifier = nn.Linear(config.hidden_size, config.num_labels)
+
+        self.init_weights()
+
+        # Model parallel
+        self.model_parallel = False
+        self.device_map = None
+
+    @add_start_docstrings_to_model_forward(GPT2_INPUTS_DOCSTRING)
+    @add_code_sample_docstrings(
+        tokenizer_class=_TOKENIZER_FOR_DOC,
+        checkpoint="microsoft/DialogRPT-updown",
+        output_type=TokenClassifierOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    def forward(
+        self,
+        input_ids=None,
+        past_key_values=None,
+        attention_mask=None,
+        token_type_ids=None,
+        position_ids=None,
+        head_mask=None,
+        inputs_embeds=None,
+        labels=None,
+        use_cache=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+    ):
+        r"""
+        labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size,)`, `optional`):
+            Labels for computing the sequence classification/regression loss. Indices should be in :obj:`[0, ...,
+            config.num_labels - 1]`. If :obj:`config.num_labels == 1` a regression loss is computed (Mean-Square loss),
+            If :obj:`config.num_labels > 1` a classification loss is computed (Cross-Entropy).
+        """
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        transformer_outputs = self.transformer(
+            input_ids,
+            past_key_values=past_key_values,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        hidden_states = transformer_outputs[0]
+        hidden_states = self.dropout(hidden_states)
+        logits = self.classifier(hidden_states)
+
+        loss = None
+        if labels is not None:
+            loss_fct = CrossEntropyLoss()
+            # Only keep active parts of the loss
+            if attention_mask is not None:
+                active_loss = attention_mask.view(-1) == 1
+                active_logits = logits.view(-1, self.num_labels)
+                active_labels = torch.where(
+                    active_loss, labels.view(-1), torch.tensor(loss_fct.ignore_index).type_as(labels)
+                )
+                loss = loss_fct(active_logits, active_labels)
+            else:
+                loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
+
+        if not return_dict:
+            output = (logits,) + transformer_outputs[2:]
+            return ((loss,) + output) if loss is not None else output
+
+        return TokenClassifierOutput(
+            loss=loss,
+            logits=logits,
+            hidden_states=transformer_outputs.hidden_states,
+            attentions=transformer_outputs.attentions,
+        )