gsmyrnis commited on
Commit
b7ecb1e
·
verified ·
1 Parent(s): 1b96c8c

Upload configs.yaml with huggingface_hub

Browse files
Files changed (1) hide show
  1. configs.yaml +41 -0
configs.yaml ADDED
@@ -0,0 +1,41 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ assistant_tag: gpt
2
+ bf16: true
3
+ content_tag: value
4
+ cutoff_len: 16384
5
+ dataloader_num_workers: 4
6
+ dataloader_persistent_workers: true
7
+ dataloader_pin_memory: true
8
+ dataset: mlfoundations-dev/a1_code_code_contests
9
+ dataset_dir: ONLINE
10
+ ddp_timeout: 180000000
11
+ deepspeed: dcft/train/zero3.json
12
+ do_train: true
13
+ enable_liger_kernel: false
14
+ eval_strategy: 'no'
15
+ finetuning_type: full
16
+ formatting: sharegpt
17
+ global_batch_size: 128
18
+ gradient_accumulation_steps: 4
19
+ hub_model_id: mlfoundations-dev/a1_code_code_contests
20
+ include_hp: dcft/train/hp_settings/paper/reasoning_medium.yaml
21
+ learning_rate: 4.0e-05
22
+ logging_steps: 1
23
+ lr_scheduler_type: cosine
24
+ messages: conversations
25
+ model_name_or_path: Qwen/Qwen2.5-7B-Instruct
26
+ num_train_epochs: 5.0
27
+ output_dir: /tmp/dcft_checkpoints/train/checkpoints/a1_code_code_contests
28
+ overwrite_cache: true
29
+ per_device_train_batch_size: 1
30
+ plot_loss: true
31
+ preprocessing_num_workers: 4
32
+ push_to_db: true
33
+ push_to_hub: true
34
+ report_to: wandb
35
+ role_tag: from
36
+ run_name: a1_code_code_contests
37
+ save_strategy: epoch
38
+ stage: sft
39
+ template: qwen25
40
+ user_tag: human
41
+ warmup_ratio: 0.1