BioReason-0813/run.sh

echo "Starting GRPO training..."

#!/bin/bash
# run_blip2.sh
# 用于启动 BLIP2 + GRPO 训练的脚本

# ===== 基本路径配置 =====
DATA_FILE=/oss/wangyujia/ProtT3/ProtT3/data/sft/dataset/DeepLocBinary/test.csv
DATASET_NAME=deeplocbinary
OUTPUT_DIR=./output
CACHE_DIR=./cache

# ===== 模型配置 =====
BERT_PATH=/nas/shared/kilab/wangyujia/ProtT3/plm_model/microsoft
PLM_MODEL=/nas/shared/kilab/wangyujia/ProtT3/plm_model/esm2-150m
LLM_MODEL=/oss/wangyujia/BIO/construction_finetuning/alpaca/v1-20250609-141541/checkpoint-50-merged
SFT_CHECKPOINT=/nas/shared/kilab/wangyujia/ProtT3/all_checkpoints/stage2_07301646_2datasets_construct/epoch=09.ckpt/converted.ckpt
# ===== 训练参数 =====
BATCH_SIZE=4
EPOCHS=3
LR=1e-5

# ===== 奖励函数权重 =====
FORMAT_WEIGHT=0.2
ACCURACY_WEIGHT=0.6
REPETITION_WEIGHT=0.2

# ===== 运行训练 =====
python blips_reason.py \
    --data_file_paths ${DATA_FILE} \
    --dataset_name ${DATASET_NAME} \
    --reward_funcs combined \
    --format_weight ${FORMAT_WEIGHT} \
    --accuracy_weight ${ACCURACY_WEIGHT} \
    --repetition_weight ${REPETITION_WEIGHT} \
    --use_custom_prompts \
    --template_name classification \
    --max_seq_length 1000 \
    --output_dir ${OUTPUT_DIR} \
    --per_device_train_batch_size ${BATCH_SIZE} \
    --num_train_epochs ${EPOCHS} \
    --learning_rate ${LR} \
    --bert_name ${BERT_PATH} \
    --plm_model ${PLM_MODEL} \
    --llm_name ${LLM_MODEL} \
    --sft_checkpoint ${SFT_CHECKPOINT} \
    --plm_tune freeze \
    --llm_tune lora \
    --qformer_tune train \
    --lora_r 8 \
    --lora_alpha 16 \
    --lora_dropout 0.1 \
    --enable_flash \
    --cache_dir ${CACHE_DIR}



# python protein_reason.py \
#     --output_dir "./grpo_outputs" \
#     --model_name_or_path "Qwen/Qwen3-0.6B" \
#     --protein_model_name_or_path "facebook/esm2_t6_8M_UR50D" \
#     --qformer_model_name_or_path "microsoft/BiomedNLP-BiomedBERT-base-uncased-abstract-fulltext" \
#     --dataset_name "wanglab/protein_function" \
#     --sft_checkpoint "./checkpoints/best_model" \
#     --per_device_train_batch_size 4 \
#     --gradient_accumulation_steps 4 \
#     --num_train_epochs 3 \
#     --learning_rate 1e-6 \
#     --beta 0.04 \
#     --temperature 0.6 \
#     --top_p 0.95 \
#     --top_k 20 \
#     --max_completion_length 800 \
#     --num_generations 8 \
#     --reward_funcs "xmlcount" "soft_format" "strict_format" "correctness" \
#     --lora_r 32 \
#     --lora_alpha 64 \
#     --lora_dropout 0.05 \
#     --freeze_protein_modules \
#     --logging_steps 2 \
#     --eval_strategy "steps" \
#     --eval_steps 100 \
#     --save_steps 200 \
#     --report_to "wandb" \
#     --log_completions
    
# python blip2_reason.py \
#     --data_file_paths /oss/wangyujia/ProtT3/ProtT3/data/sft/dataset/DeepLocBinary/test.csv \
#     --reward_funcs combined \
#     --format_weight 0.2 \
#     --accuracy_weight 0.6 \
#     --repetition_weight 0.2 \
#     --use_custom_prompts \
#     --template_name classification \
#     --max_seq_length 1000 \
#     --output_dir ./output \
#     --per_device_train_batch_size 4 \
#     --num_train_epochs 3 \
#     --learning_rate 1e-5

echo "GRPO training completed!"

echo "All training stages completed successfully!"