run_train.sh

#!/bin/bash
# ASR 诗词纠错模型训练脚本
# 基于 ChineseErrorCorrector3-4B 继续 SFT

cd ./asr

# ============= 环境配置 =============
export CUDA_VISIBLE_DEVICES=0,1,2,3  # 根据实际 GPU 数量修改
export NCCL_P2P_DISABLE=1  # 如果遇到 NCCL 问题可取消注释

# ============= 数据路径（优先 RL 最新 v4-lite，缺失时回退 v4/v3） =============
RL_DIR="./asr/log1"
LATEST_RL_RUN=""
for d in "${RL_DIR}"/asr_v4_lite_*; do
    [ -d "${d}" ] || continue
    if [ -z "${LATEST_RL_RUN}" ] || [ "${d}" -nt "${LATEST_RL_RUN}" ]; then
        LATEST_RL_RUN="${d}"
    fi
done

if [ -n "${LATEST_RL_RUN}" ] && [ -f "${LATEST_RL_RUN}/train.jsonl" ] && [ -f "${LATEST_RL_RUN}/valid.jsonl" ]; then
    TRAIN_FILE="${LATEST_RL_RUN}/train.jsonl"
    DEV_FILE="${LATEST_RL_RUN}/valid.jsonl"
elif [ -f "train_data_v4/train.jsonl" ]; then
    echo "WARNING: RL v4-lite data not found, falling back to v4"
    TRAIN_FILE="train_data_v4/train.jsonl"
    DEV_FILE="train_data_v4/valid.jsonl"
else
    echo "WARNING: v4 data not found, falling back to v3"
    TRAIN_FILE="train_data_v3/train.jsonl"
    DEV_FILE="train_data_v3/valid.jsonl"
fi

echo "Using train file: ${TRAIN_FILE}"
echo "Using dev file:   ${DEV_FILE}"

# ============= 单卡训练 =============
# 显存 < 24GB 建议使用 QLoRA (int4=True, qlora=True)
python train_asr_poetry_correction.py \
    --model_name ChineseErrorCorrector3-4B \
    --train_file ${TRAIN_FILE} \
    --dev_file ${DEV_FILE} \
    --output_dir output/asr_poetry_lora \
    --num_train_epochs 3 \
    --learning_rate 2e-5 \
    --per_device_train_batch_size 4 \
    --gradient_accumulation_steps 4 \
    --lora_r 16 \
    --lora_alpha 32 \
    --save_steps 500 \
    --eval_steps 500 \
    --early_stopping_patience 2 \
    --correction_token_weight 4.0 \
    --logging_steps 50 \
    --bf16 True

# ============= 多卡训练 (推荐) =============
# 取消下面的注释使用多卡训练
# torchrun --nproc_per_node=4 train_asr_poetry_correction.py \
#     --model_name ChineseErrorCorrector3-4B \
#     --train_file ${TRAIN_FILE} \
#     --dev_file ${DEV_FILE} \
#     --output_dir output/asr_poetry_lora \
#     --num_train_epochs 3 \
#     --learning_rate 2e-5 \
#     --per_device_train_batch_size 4 \
#     --gradient_accumulation_steps 1 \
#     --lora_r 16 \
#     --lora_alpha 32 \
#     --save_steps 500 \
#     --eval_steps 500 \
#     --logging_steps 50 \
#     --bf16 True

# ============= QLoRA 训练 (省显存) =============
# 如果显存不足，使用 QLoRA:
# python train_asr_poetry_correction.py \
#     --model_name ChineseErrorCorrector3-4B \
#     --train_file ${TRAIN_FILE} \
#     --dev_file ${DEV_FILE} \
#     --output_dir output/asr_poetry_qlora \
#     --int4 True \
#     --qlora True \
#     --num_train_epochs 3 \
#     --per_device_train_batch_size 8 \
#     --gradient_accumulation_steps 2 \
#     --lora_r 16 \
#     --lora_alpha 32

echo "训练完成！"