Continue FT from stage 2 with custom data

[adrielkuek]

Hi, was wondering whether the stage2 script would be applicable for further FT from stage 2 with a small custom dataset for domain transfer? Or do we have to rewrite a separate script to do this?

Thanks and appreciate any help given!

Regards,

Adriel

[yanwei-li]

Hi, thanks for your interest. You can try this, and replace the data to your custom dataset

FINETUNE_NAME=Mini-Gemini-7B
STAGE3_NAME=Your_prefer_name
AUX_SIZE=768
deepspeed minigemini/train/train_mem.py \
    --deepspeed ./scripts/zero2_offload.json \
    --model_name_or_path ./work_dirs/$FINETUNE_NAME \
    --version v1 \
    --data_path ./data/MiniGemini-Finetune/minigemini_instruction.json \
    --image_folder ./data/MiniGemini-Finetune \
    --vision_tower model_zoo/OpenAI/clip-vit-large-patch14-336 \
    --vision_tower_aux model_zoo/OpenAI/openclip-convnext-large-d-320-laion2B-s29B-b131K-ft-soup \
    --mm_projector_type mlp2x_gelu \
    --mm_vision_select_layer -2 \
    --mm_use_im_start_end False \
    --mm_use_im_patch_token False \
    --image_aspect_ratio pad \
    --group_by_modality_length True \
    --image_size_aux $AUX_SIZE \
    --bf16 True \
    --output_dir ./work_dirs/$STAGE3_NAME \
    --num_train_epochs 1 \
    --per_device_train_batch_size 8 \
    --per_device_eval_batch_size 4 \
    --gradient_accumulation_steps 2 \
    --evaluation_strategy "no" \
    --save_strategy "steps" \
    --save_steps 1000 \
    --save_total_limit 1 \
    --learning_rate 2e-5 \
    --weight_decay 0. \
    --warmup_ratio 0.03 \
    --lr_scheduler_type "cosine" \
    --logging_steps 1 \
    --tf32 True \
    --model_max_length 2048 \
    --gradient_checkpointing True \
    --dataloader_num_workers 4 \
    --lazy_preprocess True \
    --report_to wandb

[adrielkuek]

Hi Yanwei, apologies for the late response. Thank you very much for providing the guidance on this. Appreciate it!