TRAIN_DATA="/mnt/dhwfile/raise/user/panjiabao/huxiaobin/zhanghaobin/0416upload/lora_0323_10w+55w+error+syn_with_domain_train90_targeted_rl_train90_loramerged_basewer_3suppress_server.jsonl"
VAL_DATA="/mnt/dhwfile/raise/user/panjiabao/huxiaobin/zhanghaobin/0416upload/lora_0323_10w+55w+error+syn_with_domain_train90_targeted_rl_val5_sample5p_server.jsonl"   # <- 你新增的验证集

export MASTER_PORT=29540
export MASTER_ADDR=127.0.0.1

######################
# 0. 基础环境变量 (wandb)
######################
export WANDB_BASE_URL="https://api.wandb.ai"
export WANDB_API_KEY="af28909dd03c6220f0ad4df6bd9a0a1fbddb3e55"
export WANDB_PROJECT="qwen3_asr_swift_dapo"    # 对应截图里的项目名
export WANDB_ENTITY="pang_kaiyu-none"        # 对应截图里的 Entity

# 让 wandb 在多卡训练时只开一个进程写日志（可选）
export WANDB_MODE=online
export NPROC_PER_NODE=4
export SWIFT_SINGLE_DEVICE_MODE=1
export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True


export ASR_REWARD_DEBUG=1
export ASR_REWARD_DEBUG_PATH=/mnt/dhwfile/raise/user/panjiabao/huxiaobin/zhanghaobin/0416upload/out/qwen3asr_dapo_reward56_4x4x12_12gen_4GPU/reward_debug
export ASR_REWARD_DEBUG_MAX_ROWS=1000



swift rlhf \
  --rlhf_type grpo \
  --external_plugins /mnt/dhwfile/raise/user/panjiabao/huxiaobin/zhanghaobin/0416upload/my_qwen3_asr_dapo_register.py /mnt/dhwfile/raise/user/panjiabao/huxiaobin/zhanghaobin/0416upload/0417_reward.py \
  --model /mnt/dhwfile/raise/user/panjiabao/huxiaobin/zhanghaobin/qwen3-asr-merged \
  --model_type my_qwen3_asr_rl \
  --template my_qwen3_asr_rl \
  --dataset ${TRAIN_DATA} \
  --val_dataset ${VAL_DATA} \
  --reward_funcs asr_wer_sub_len_cmp_hallu_dirty_v56 \
  --train_type lora \
  --use_vllm false \
  --log_completions true \
  --loss_type dapo \
  --advantage_estimator grpo \
  --scale_rewards group \
  --num_iterations 2 \
  --beta 0.04 \
  --epsilon_high 0.28 \
  --dynamic_sample true \
  --max_resample_times 4 \
  --overlong_filter true \
  --truncation_strategy delete \
  --num_generations 12 \
  --generation_batch_size 48 \
  --per_device_train_batch_size 4 \
  --per_device_eval_batch_size 4 \
  --gradient_accumulation_steps 12 \
  --num_generations_eval 4 \
  --max_completion_length 256 \
  --temperature 0.50 \
  --top_p 0.95 \
  --top_k 50 \
  --repetition_penalty 1.08 \
  --learning_rate 5e-5 \
  --lr_scheduler_type cosine \
  --warmup_ratio 0.03 \
  --report_to wandb \
  --run_name qwen3asr_dapo_reward56_4x4x12_12gen_4GPU \
  --output_dir /mnt/dhwfile/raise/user/panjiabao/huxiaobin/zhanghaobin/0416upload/out/qwen3asr_dapo_reward56_4x4x12_12gen_4GPU \
  --save_strategy steps \
  --save_steps 20 \
  --logging_steps 5 \
  --freeze_llm false \
  --freeze_vit false \
  --freeze_aligner false \
  --remove_unused_columns false \
  --padding_side left