MLLM Finetuning Demo

MLLM-Finetuning-Demo

Install LLaMA-Factory

git clone https://github.com/hiyouga/LLaMA-Factory.git
cd LLaMA-Factory
pip install -e .[torch,metrics]
cd .. # 回到项目根目录

Pre-training

Feature alignment in LLaVA, freeze language_model and vision_tower , and only fine-tune multi_modal_projector .

CUDA_VISIBLE_DEVICES=0 llamafactory-cli train config/llava_pt.yaml

Fine adjustment

CUDA_VISIBLE_DEVICES=0 llamafactory-cli train config/llava_lora_sft.yaml

Web Chat

CUDA_VISIBLE_DEVICES=0 llamafactory-cli webchat 
--model_name_or_path llava-hf/llava-1.5-7b-hf 
--adapter_name_or_path saves/llava1_5-7b/lora/sft 
--template llava

Upload dataset to Huggingface

Please replace your own key in upload_dataset.py .

python3 upload_dataset.py

Export and upload models to huggingface

Please replace your own export_hub_model_id and hf_hub_token in config/llava_lora_sft_export.yaml .

CUDA_VISIBLE_DEVICES=0 llamafactory-cli export config/llava_lora_sft_export.yaml