Can't Reproduce the result of the Bunny bsaed Phi-3

[bollossom]

I have used the bunny_695k for finetune with unfreeze sigclip, however I found that science QA only get 68.3%.
Is this because the dataset is too small during fine-tuning and the weights of Vision Tower should not be unfrozen?

Traing script:

deepspeed bunny/train/train.py \
    --lora_enable True --lora_r 128 --lora_alpha 256 --mm_projector_lr 2e-5 \
    --deepspeed ./script/deepspeed/zero3.json \
    --model_name_or_path .LLaVA/llms/Phi_3_mini_4k \
    --model_type $MODEL_TYPE \
    --version phi3 \
    --data_path ./finetune/bunny_695k.json \
    --image_folder .bunny/finetune/images \
    --vision_tower `./LLaVA/vision_tower/siglip_L_384` \
    --use_s2 True \
    --unfreeze_vision_tower True \
    --pretrain_mm_mlp_adapter ./checkpoints-pretrain/$PRETRAIN_DIR/mm_projector.bin \
    --mm_projector_type mlp2x_gelu \
    --image_aspect_ratio pad \
    --group_by_modality_length False \
    --bf16 True \
    --output_dir ./checkpoints-$MODEL_TYPE/$OUTPUT_DIR \
    --num_train_epochs 1 \
    --per_device_train_batch_size 4 \
    --per_device_eval_batch_size 4 \
    --gradient_accumulation_steps 4 \
    --evaluation_strategy "no" \
    --save_strategy "steps" \
    --save_steps 500 \
    --save_total_limit 1 \
    --learning_rate 2e-4 \
    --weight_decay 0. \
    --warmup_ratio 0.03 \
    --lr_scheduler_type "cosine" \
    --logging_steps 1 \
    --tf32 True \
    --model_max_length 4096 \
    --gradient_checkpointing True \
    --dataloader_num_workers 4 \
    --lazy_preprocess True \
    --run_name bunny_phi3_finetune \
    --report_to wandb | tee 2>&1 ./checkpoints-$MODEL_TYPE/$OUTPUT_DIR/log.txt

[Isaachhh]

Bunny uses SigLIP-SO.

Besides, you may refer to the paper and related results of Bunny-v1.0-4B.