hulehule/pandada

这是Q4个性化新项目PPD：Personalized prompt DPO

每次更新完代码一定注意git提交！

1. 安装依赖 pip install -r requirements.txt

2. 下载数据 python dataset/lamp_downloader.py --task 4 --task 5 --task 7 python dataset/longlamp_downloader.py

3. 用BGE生成embedding

python3 - << 'EOF' from huggingface_hub import snapshot_download

snapshot_download( repo_id="BAAI/bge-base-en-v1.5", local_dir="./bge-base-en-v1.5", local_dir_use_symlinks=False ) EOF

CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python embedding.py

3. 运行GPU压力测试 python gpu_utils/gpu_stress.py --target_utilization 0.5 --num_gpus 4

CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python /opt/tiger/PPD/inference/RAG_bm25_qwen_lamp5_concise.py
--questions_file /opt/tiger/PPD/data/lamp/LaMP_5/user/dev/dev_questions.json
--outputs_file /opt/tiger/PPD/data/lamp/LaMP_5/user/dev/dev_outputs.json
--output_file /opt/tiger/PPD/data/lamp/LaMP_5/user/dev/predictions_qwen3_zero_shot_concise.json
--model_name Qwen/Qwen3-8B
--device cuda
--torch_dtype bfloat16
--max_new_tokens 128
--batch_size 64

CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python /opt/tiger/PPD/inference/zero_shot_qwen_lamp5_concise.py
--questions_file /opt/tiger/PPD/data/lamp/LaMP_5/user/dev/dev_questions.json
--outputs_file /opt/tiger/PPD/data/lamp/LaMP_5/user/dev/dev_outputs.json
--output_file /opt/tiger/PPD/data/lamp/LaMP_5/user/dev/zero_shot_qwen3_concise.json
--model_name Qwen/Qwen3-8B
--device cuda
--torch_dtype bfloat16
--max_new_tokens 128
--batch_size 64

CUDA_VISIBLE_DEVICES= 7 python /opt/tiger/PPD/inference/RAG_bm25_qwen_lamp5_concise.py
--questions_file /opt/tiger/PPD/data/lamp/LaMP_5/user/train/train_questions.json
--outputs_file /opt/tiger/PPD/data/lamp/LaMP_5/user/train/train_outputs.json
--output_file /opt/tiger/PPD/data/lamp/LaMP_5/user/dev/RAG_qwen3_concise.json
--model_name Qwen/Qwen3-8B
--device cuda
--torch_dtype bfloat16
--max_new_tokens 128
--batch_size 64

/opt/tiger/PPD/data/lamp/LaMP_5/user/dev/predictions_qwen3_zero_shot_comdy.json

python3 -c "import nltk; nltk.download('wordnet'); nltk.download('omw-1.4')"

CUDA_VISIBLE_DEVICES=0,1,2,3
accelerate launch
--num_processes 4
--mixed_precision bf16
--main_process_port 0
/opt/tiger/PPD/DPO/train_dpo_qwen3_lamp5_gt_vs_pred.py
--model_name Qwen/Qwen3-8B
--cleaned_file ./cleaned_predictions.json
--output_dir ./qwen3_lamp5_dpo_gt_vs_pred
--per_device_train_batch_size 1
--gradient_accumulation_steps 16
--learning_rate 5e-6
--num_train_epochs 1
--beta 0.1
--max_length 128
--max_prompt_length 256

/opt/tiger/PPD/data/lamp/LaMP_5/user/dev/predictions_qwen3_zero_shot_comdy.json

python clean_and_score.py
--input /opt/tiger/PPD/data/lamp/LaMP_5/user/dev/predictions_qwen3_zero_shot_direct.json
--output cleaned_predictions_predictions_qwen3_zero_shot.json

CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python /opt/tiger/PPD/inference/LaMP_2/zero_shot_qwen_lamp2.py --questions_file /opt/tiger/PPD/data/lamp/LaMP_2/user/dev/dev_questions.json --outputs_file /opt/tiger/PPD/data/lamp/LaMP_2/user/dev/dev_outputs.json --output_file /opt/tiger/PPD/data/lamp/LaMP_2/user/dev/qwen3_zero_shot_chat.json --batch_size 64

CUDA_VISIBLE_DEVICES=1,2,3,4,5,6,7 python 10hours.py