forked from shibing624/MedicalGPT
-
Notifications
You must be signed in to change notification settings - Fork 0
/
run_ppo.sh
26 lines (26 loc) · 830 Bytes
/
run_ppo.sh
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
CUDA_VISIBLE_DEVICES=0,1 torchrun --nproc_per_node 2 ppo_training.py \
--model_type auto \
--model_name_or_path Qwen/Qwen1.5-0.5B-Chat \
--reward_model_name_or_path OpenAssistant/reward-model-deberta-v3-large-v2 \
--template_name qwen \
--torch_dtype float16 \
--device_map auto \
--train_file_dir ./data/finetune \
--validation_file_dir ./data/finetune \
--batch_size 8 \
--max_source_length 1024 \
--max_target_length 256 \
--max_train_samples 1000 \
--use_peft True \
--target_modules q_proj,v_proj \
--lora_rank 8 \
--lora_alpha 16 \
--lora_dropout 0.05 \
--do_train \
--max_steps 100 \
--learning_rate 1e-5 \
--save_steps 50 \
--output_dir outputs-rl-qwen-v1 \
--early_stopping True \
--target_kl 0.1 \
--reward_baseline 0.0