MedCCO

This repository provides the official implementation of MedCCO.

---

🔧 Installation

Please refer to the verl folder for detailed installation instructions.

---

📚 Datasets

We use publicly available datasets for training and evaluation. Below are the links for download:

• VQA-RAD
• SLAKE
• PathVQA
• Quilt-VQA
• MedXpertQA
• For PMC-VQA, OmniMedVQA, and MMMU Health & Medical track, we use the test set from Medical_Multimodal_Evaluation_Data.

For any dataset-related issues, feel free to contact us.

---

🧹 Dataset Preprocessing

We refine the open-ended VQA consistency in VQA-RAD, SLAKE, and PathVQA, as detailed in our paper.

Requirements:

• At least 2 GPUs with 80GB VRAM for Qwen2.5-VL-72B
• We use vLLM to accelerate inference

To preprocess:

cd preprocess
python clean_medvqa.py

GRPO Training Pipeline

• Step 1: Close-ended QA training

cd verl
bash examples/train_medcco/train_close_qwen2_5_vl-7b.sh

• Step 2: Open-ended QA training

cd verl
bash examples/train_medcco/train_open_qwen2_5_vl-7b.sh

Inference

• Step 1: Deploy the trained model with vLLM

cd inference
bash deploy_vllm.sh

• Step 2: Run inference on test datasets

cd inference
python eval_vllm_verl_models.py

Our Models

Due to anonymity requirements during review, the model checkpoint link is currently left blank. If requested by reviewers, we will provide access to the weights.

• MedCCO-7B
• MedCCO-3B

Our Datasets

Due to anonymity requirements during review, the dataset link is currently left blank. If requested by reviewers, we will provide access to the datasets.

• MedCCO-TrainingDatasets