LLaVa3-Med

We apply 3-stages to train our model.

1. Pretraining: We utilize a dataset comprising 600k image-text pairs from PMC and 60k medical references based on Mayo Clinic guidelines for the pretraining phase.
2. Instruction Fine-tuning: We employ a dataset consisting of 60k LLaVA_Med instruction fine-tuning examples and PMC-VQA datasets to perform instruction learning.
3. Fine-tuning: Our model undergoes fine-tuning on various VQA datasets.

Inference

CUDA_VISIBLE_DEVICES=0 python -m evaluation \
        --model-path model_path \
        --question-file data_path \
        --image-folder image_path \
        --answers-file result.jsonl \
        --temperature 0.7 \
        --conv-mode llama3

Results

Because GPT-4 has not been fine-tuned on these VQA tasks, the answers it generates for open questions differ significantly in style from the reference answers. Therefore, we employed a few-shot approach and modified GPT-4's answers to match the style of the reference answers.

Dataset	Metric	Med-Gemini	Med-PaLM-540B	GPT-4V	LLaVa3-Med
Slake-VQA	Token F1	87.5	89.3	76.8	89.8†
Path-VQA	Token F1	64.7	62.7	57.7	64.9†

Table 1 | Multimodal evaluation. Performance comparison of LLaVa3-Med versus state-of-the-art (SoTA) methods.