peacock-data-public-datasets-idc-bigscience
/
train
/tr5-1B3-multilingual
/generate_dataset_probabilities.py
import json | |
import os | |
from pathlib import Path | |
def removeprefix(string, prefix): | |
if string.startswith(prefix): | |
string = string[len(prefix):] | |
return string | |
def removesuffix(string, suffix): | |
if string.endswith(suffix): | |
string = string[:-len(suffix)] | |
return string | |
def convert_path(original_path, data_path): | |
prefix_to_remove = "dumped/mc4_processed_data/" | |
suffix_to_remove = ".bin" | |
return data_path / removesuffix(removeprefix(original_path, prefix_to_remove), suffix_to_remove) | |
def main(): | |
"""Write to """ | |
data_path = Path(os.environ["six_ALL_CCFRSCRATCH"]) / "datasets-custom" / "mc4" / "mc4_preprocessing" | |
output_path = Path(os.environ["six_ALL_CCFRSCRATCH"]) / "checkpoints" / "tr5-1B3-multilingual" / "dataset_probabilities.txt" | |
probabilies_path = data_path / "sample_iterator_probs" / "iterator_selection_prob.0.3.train.json" | |
with open(probabilies_path, "r") as fprob: | |
probabilities = json.load(fprob) | |
# Format probabilities dictionary to store path in key and probability as value | |
probabilities = { | |
convert_path(key, data_path): value[0] for key, value in probabilities.items() | |
} | |
with open(output_path, "w") as fout: | |
fout.write(" ".join([f"{prob} {path}" for path, prob in probabilities.items()])) | |
pass | |
if __name__ == "__main__": | |
main() | |