lzw-lzw
commented
7 months ago 我也遇到了这个问题，请问你解决了吗？
Reminder
- [x] I have read the README and searched the existing issues.
Reproduction
cd LLaMA-Factory && HF_ENDPOINT=https://hf-mirror.com accelerate launch src/train_bash.py \ --stage sft \ --do_train \ --model_name_or_path codellama/CodeLlama-7b-Python-hf \ --dataset codealpaca,codeforces_python_submissions_sft \ --template default \ --finetuning_type lora \ --lora_target q_proj,v_proj \ --output_dir output/sft/test_train \ --overwrite_cache \ --per_device_train_batch_size 4 \ --gradient_accumulation_steps 4 \ --lr_scheduler_type cosine \ --logging_steps 10 \ --save_steps 500 \ --learning_rate 5e-5 \ --num_train_epochs 3.0 \ --plot_loss \ --fp16 cd LLaMA-Factory && HF_DATASETS_OFFLINE=1 TRANSFORMERS_OFFLINE=1 accelerate launch src/train_bash.py \ --stage rm \ --do_train \ --model_name_or_path codellama/CodeLlama-7b-Python-hf \ --adapter_name_or_path output/sft/test_train \ --create_new_adapter \ --dataset codeforces_python_submissions_rl \ --template default \ --finetuning_type lora \ --lora_target q_proj,v_proj \ --output_dir output/rm/test_train \ --per_device_train_batch_size 1 \ --gradient_accumulation_steps 16 \ --lr_scheduler_type cosine \ --logging_steps 10 \ --save_steps 500 \ --learning_rate 1e-4 \ --num_train_epochs 1.0 \ --plot_loss \ --fp16 cd LLaMA-Factory && HF_DATASETS_OFFLINE=1 TRANSFORMERS_OFFLINE=1 accelerate launch src/train_bash.py \ --stage ppo \ --do_train \ --model_name_or_path codellama/CodeLlama-7b-Python-hf \ --adapter_name_or_path output/sft/test_train \ --create_new_adapter \ --dataset codealpaca,codeforces_python_submissions_sft \ --template default \ --finetuning_type lora \ --lora_target q_proj,v_proj \ --reward_model output/rm/test_train \ --output_dir output/ppo/test_train \ --per_device_train_batch_size 1 \ --gradient_accumulation_steps 16 \ --lr_scheduler_type cosine \ --top_k 0 \ --top_p 0.8 \ --logging_steps 10 \ --save_steps 100 \ --learning_rate 1e-5 \ --num_train_epochs 1.0 \ --plot_loss \ --fp16Expected behavior
在PPO训练阶段，突然出现loss突刺，reward也变成了负数，这有可能是啥原因呢？
System Info
No response
Others
No response
koking0
cabisarri
Reminder
Reproduction
Expected behavior
在PPO训练阶段，突然出现loss突刺，reward也变成了负数，这有可能是啥原因呢？
System Info
No response
Others
No response