🩺 MedEvalKit: A Unified Medical Evaluation Framework

📖 arXiv Paper • 🤗 Lingshu Models • 🌐 Lingshu Project Page

---

📌 Introduction

A comprehensive evaluation framework for Large Medical Models (LMMs/LLMs) in the healthcare domain.
We welcome contributions of new models, benchmarks, or enhanced evaluation metrics!

---

Eval Results

MultiModal Eval Results

Text-Only Eval Results

🔥 Latest News

• 2025-06-12 - Initial release of MedEvalKit v1.0!

---

🧪 Supported Benchmarks

Multimodal Medical Benchmarks	Text-Only Medical Benchmarks
MMMU-Medical-test	MedQA-USMLE
MMMU-Medical-val	MedMCQA
PMC_VQA	PubMedQA
OmniMedVQA	Medbullets-op4
IU XRAY	Medbullets-op5
MedXpertQA-Multimodal	MedXpertQA-Text
CheXpert Plus	SuperGPQA
MIMIC-CXR	HealthBench
VQA-RAD	CMB
SLAKE	CMExam
PATH-VQA	CMMLU
MedFrameQA	MedQA-MCMLE

---

🤖 Supported Models

HuggingFace Exclusive

• BiMediX2
• BiomedGPT
• HealthGPT
• Janus
• Med_Flamingo
• MedDr
• MedGemma
• NVILA
• VILA_M3

HF + vLLM Compatible

• HuatuoGPT-vision
• InternVL
• Llama_3.2-vision
• LLava
• LLava_Med
• Qwen2_5_VL
• Qwen2_VL

---

🛠️ Installation

# Clone repository
git clone https://github.com/DAMO-NLP-SG/MedEvalKit
cd MedEvalKit

# Install dependencies
pip install -r requirements.txt
pip install 'open_clip_torch[training]'
pip install flash-attn --no-build-isolation

# For LLaVA-like models
git clone https://github.com/LLaVA-VL/LLaVA-NeXT.git
cd LLaVA-NeXT && pip install -e .

---

📂 Dataset Preparation

HuggingFace Datasets (Direct Access)

# Set DATASETS_PATH='hf'
VQA-RAD: flaviagiammarino/vqa-rad
SuperGPQA: m-a-p/SuperGPQA
PubMedQA: openlifescienceai/pubmedqa
PATHVQA: flaviagiammarino/path-vqa
MMMU: MMMU/MMMU
MedQA-USMLE: GBaker/MedQA-USMLE-4-options
MedQA-MCMLE: shuyuej/MedQA-MCMLE-Benchmark
Medbullets_op4: tuenguyen/Medical-Eval-MedBullets_op4
Medbullets_op5: LangAGI-Lab/medbullets_op5
CMMMU: haonan-li/cmmlu
CMExam: fzkuji/CMExam
CMB: FreedomIntelligence/CMB
MedFrameQA: SuhaoYu1020/MedFrameQA

Local Datasets (Manual Download Required)

Dataset	Source
MedXpertQA	TsinghuaC3I
SLAKE	BoKelvin
PMCVQA	RadGenome
OmniMedVQA	foreverbeliever
MIMIC_CXR	MIMIC_CXR
IU_Xray	IU_Xray
CheXpert Plus	CheXpert Plus
HealthBench	Normal,Hard,Consensus

---

🚀 Quick Start

1. Configure eval.sh

#!/bin/bash
export HF_ENDPOINT=https://hf-mirror.com
# MMMU-Medical-test,MMMU-Medical-val,PMC_VQA,MedQA_USMLE,MedMCQA,PubMedQA,OmniMedVQA,Medbullets_op4,Medbullets_op5,MedXpertQA-Text,MedXpertQA-MM,SuperGPQA,HealthBench,IU_XRAY,CheXpert_Plus,MIMIC_CXR,CMB,CMExam,CMMLU,MedQA_MCMLE,VQA_RAD,SLAKE,PATH_VQA,MedFrameQA
EVAL_DATASETS="Medbullets_op4" 
DATASETS_PATH="hf"
OUTPUT_PATH="eval_results/{}"
# TestModel,Qwen2-VL,Qwen2.5-VL,BiMediX2,LLava_Med,Huatuo,InternVL,Llama-3.2,LLava,Janus,HealthGPT,BiomedGPT,Vllm_Text,MedGemma,Med_Flamingo,MedDr
MODEL_NAME="Qwen2.5-VL"
MODEL_PATH="Qwen2.5-VL-7B-Instruct"

#vllm setting
CUDA_VISIBLE_DEVICES="0"
TENSOR_PARALLEL_SIZE="1"
USE_VLLM="False"

#Eval setting
SEED=42
REASONING="False"
TEST_TIMES=1


# Eval LLM setting
MAX_NEW_TOKENS=8192
MAX_IMAGE_NUM=6
TEMPERATURE=0
TOP_P=0.0001
REPETITION_PENALTY=1

# LLM judge setting
USE_LLM_JUDGE="True"
# gpt api model name
GPT_MODEL="gpt-4.1-2025-04-14"
OPENAI_API_KEY=""


# pass hyperparameters and run python sccript
python eval.py \
    --eval_datasets "$EVAL_DATASETS" \
    --datasets_path "$DATASETS_PATH" \
    --output_path "$OUTPUT_PATH" \
    --model_name "$MODEL_NAME" \
    --model_path "$MODEL_PATH" \
    --seed $SEED \
    --cuda_visible_devices "$CUDA_VISIBLE_DEVICES" \
    --tensor_parallel_size "$TENSOR_PARALLEL_SIZE" \
    --use_vllm "$USE_VLLM" \
    --max_new_tokens "$MAX_NEW_TOKENS" \
    --max_image_num "$MAX_IMAGE_NUM" \
    --temperature "$TEMPERATURE"  \
    --top_p "$TOP_P" \
    --repetition_penalty "$REPETITION_PENALTY" \
    --reasoning "$REASONING" \
    --use_llm_judge "$USE_LLM_JUDGE" \
    --judge_gpt_model "$GPT_MODEL" \
    --openai_api_key "$OPENAI_API_KEY" \
    --test_times "$TEST_TIMES" 

2. Run Evaluation

chmod +x eval.sh  # Add execute permission
./eval.sh

---

📜 Citation

@article{xu2025lingshu,
  title={Lingshu: A Generalist Foundation Model for Unified Multimodal Medical Understanding and Reasoning},
  author={Xu, Weiwen and Chan, Hou Pong and Li, Long and Aljunied, Mahani and Yuan, Ruifeng and Wang, Jianyu and Xiao, Chenghao and Chen, Guizhen and Liu, Chaoqun and Li, Zhaodonghui and others},
  journal={arXiv preprint arXiv:2506.07044},
  year={2025}
}

_{Built with ❤️ by the DAMO Academy Medical AI Team}