s1 swift微调效果极差

[Jack-ctrl6]

export NPROC_PER_NODE=2
export CUDA_VISIBLE_DEVICES=0,1
export MASTER_PORT=39312
export PYTORCH_CUDA_ALLOC_CONF='expandable_segments:True'
export MAX_NUM=1

GPUS=2
BATCH_SIZE=8
PER_DEVICE_BATCH_SIZE=4
GRADIENT_ACC=$((BATCH_SIZE / PER_DEVICE_BATCH_SIZE / GPUS))
LR=0.00001
EPOCHS=20.0
SAVE_STEPS=250
MAX_LEN=8192
OUTPUR_DIR='outputmini'
DATASET_DIR='train.jsonl'
model_type='interns1'
model_path='Shanghai_AI_Laboratory/Intern-S1-mini'
dataset=''

swift sft \
    --model $model_path \
    --train_type lora \
    --dataset $DATASET_DIR \
    --freeze_vit true \
    --freeze_aligner true \
    --freeze_llm false \
    --learning_rate 1e-4 \
    --lora_rank 16 \
    --lora_alpha 32 \
    --torch_dtype bfloat16 \
    --num_train_epochs $EPOCHS \
    --gradient_checkpointing true \
    --per_device_train_batch_size $PER_DEVICE_BATCH_SIZE \
    --per_device_eval_batch_size 1 \
    --learning_rate $LR \
    --gradient_accumulation_steps $GRADIENT_ACC \
    --save_strategy steps \
    --save_steps $SAVE_STEPS \
    --split_dataset_ratio 0.01 \
    --eval_strategy steps \
    --eval_steps $SAVE_STEPS \
    --save_total_limit 100 \
    --logging_steps 1 \
    --max_length $MAX_LEN \
    --output_dir $OUTPUR_DIR \
    --warmup_ratio 0.1 \
    --dataloader_num_workers 32 \
    --dataset_num_proc 16 \
    --deepspeed zero3 \
    --report_to tensorboard \
    --use_liger_kernel true \
    --attn_impl flash_attn \
    --truncation_strategy delete \
    2>&1 | tee ${OUTPUR_DIR}/training_log.txt

lora微调s1后效果极差，连简单的ocr识别都错了