Chinese-LLaVA-Med

Benchmark

Method	llava-med-zh-eval Qwen Score
GPT4 Ground Truth	68.26
LLaVA-1.5-7B	53.13
Chinese-LLaVA-Med-7B	58.78

Demo

分析组织切片

分析胸片

分析CT

Training your own Medical MLLM

Dataset

Dataset	Description
llava-med-zh-instruct-60k	60k instruction tuning data
llava-med-zh-eval	115 evaluation data

Environment

# install LLaMA-Factory
git clone https://github.com/hiyouga/LLaMA-Factory.git
cd LLaMA-Factory
pip install -e .[torch,metrics]

Finetuning

We recommend using full finetuning, but you could also use lora yaml.

# full finetuning
CUDA_VISIBLE_DEVICES=0,1 python -m torch.distributed.run \
    --nproc_per_node 2 \
    --nnodes 1 \
    --standalone \
    ../LLaMA-Factory/src/train.py config/llava1_5_full_sft.yaml

# export
# modify your own export_hub_model_id and hf_hub_token in the config/llava1_5_full_sft_export.yaml
CUDA_VISIBLE_DEVICES=0 llamafactory-cli export config/llava1_5_full_sft_export.yaml

Evaluation

# generate output results
python3 evaluation/generate_eval_content.py --model_name_or_path models/llava1_5-7b-med

# eval by qwen-1.5-14b-chat
python3 evaluation/eval_qwen_score.py --input_path outputs/llava_med_zh_eval_llava1_5-7b-med.json

Inference

# with final model
llamafactory-cli webchat config/llava1_5_full_sft_infer.yaml