MedVLSynther/MedVLSynther-1K

Dataset

stars

commits

linked in READMEs

Oct 29, 2025

updated

Browse cluster: Medical Vision-Language Synthesis Datasets →

README

Code: https://github.com/UCSC-VLAA/MedVLSynther Project Page: https://ucsc-vlaa.github.io/MedVLSynther/

📊 Datasets

Available Datasets

We release MedSynVQA and the subsets used in our paper. Each set targets medical vision–language QA and supports RLVR/SFT training.

Dataset	Generator	Verifier	Modality	Description	Download
MedSynVQA	GLM-4.5V 108B	Qwen2.5-VL 72B	Image–Text	Full training set for medical VQA (used for RLVR).	🤗 HF
MedSynVQA-10K	GLM-4.5V 108B	Qwen2.5-VL 72B	Image–Text	10K-sample training subset for RLVR.	🤗 HF
MedSynVQA-5K	GLM-4.5V 108B	Qwen2.5-VL 72B	Image–Text	5K-sample training subset for RLVR.	🤗 HF
MedSynVQA-2K	GLM-4.5V 108B	Qwen2.5-VL 72B	Image–Text	2K-sample training subset for RLVR.	🤗 HF
MedSynVQA-1K	GLM-4.5V 108B	Qwen2.5-VL 72B	Image–Text	1K-sample training subset for RLVR.	🤗 HF
MedSynVQA-5K-qwen-glm	Qwen2.5-VL 72B	GLM-4.5V 108B	Image–Text	5K subset for generator and verifier choice ablation (GLM→Qwen generator, Qwen→GLM verifier).	🤗 HF
MedSynVQA-5K-internvl-glm	InternVL-3.5 38B	GLM-4.5V 108B	Image–Text	5K subset for generator choice ablation (InternVL→GLM verifier).	🤗 HF
MedSynVQA-5K-glm-glm	GLM-4.5V 108B	GLM-4.5V 108B	Image–Text	5K subset for verifier choice ablation (Qwen→GLM verifier).	🤗 HF
MedSynVQA-5K-no-verify	GLM-4.5V 108B	N/A	Image–Text	5K subset for verifier necessity ablation (no verification step).	🤗 HF
MedSynVQA-5K-PMC-style	GLM-4.5V 108B	N/A	Image–Text	5K subset generated with PMC-VQA–style prompts.	🤗 HF
MedSynVQA-5K-SFT	GLM-4.5V 108B	N/A	Image–Text	5K subset generated for SFT training.	🤗 HF

Dataset Usage

from datasets import load_dataset

# Load evaluation dataset
eval_dataset = load_dataset("UCSC-VLAA/MedVLThinker-Eval")

# Load training dataset
train_dataset = load_dataset("MedVLSynther/MedSynVQA-13K")

Dataset details and preparation of your own

Data Format

All train datasets follow a unified format:

{
    "images": [PIL.Image],           # List of images                           
    "question": str,                 # Question text
    "options": Dict[str, str],       # Multiple choice options
    "answer_label": str,             # Correct answer label (A, B, C, D, E)
    "answer": str,                   # Full answer text
    "reasoning": str,                # Chain-of-thought reasoning (optional)
    "dataset_name": str,             # Source dataset name
    "dataset_index": int             # Unique sample identifier
}

Contributors

NingsenWang

9 commits

Similar repos

MedVLSynther/MedVLSynther-10K

Dataset

MedVLSynther/MedVLSynther-2K

Dataset

MedVLSynther/MedVLSynther-13K

Dataset

MedVLSynther/MedVLSynther-5K

Dataset

MedVLSynther/MedVLSynther-5K-qwen-glm

Dataset

MedVLSynther/MedVLSynther-5K-internvl-glm

Dataset

MedVLSynther/MedVLSynther-5K-glm-glm

Dataset

MedVLSynther/MedVLSynther-5K-SFT

Dataset

Linked in READMEs

UCSC-VLAA/MedVLSynther

[ICLR'26] MedVLSynther: Synthesizing High-Quality Visual Question Answering from Medical Documents…

README badge

MedVLSynther/MedVLSynther-1K

Dataset

stars

commits

linked in READMEs

Oct 29, 2025

updated

Browse cluster: Medical Vision-Language Synthesis Datasets →

README

Code: https://github.com/UCSC-VLAA/MedVLSynther Project Page: https://ucsc-vlaa.github.io/MedVLSynther/

📊 Datasets

Available Datasets

We release MedSynVQA and the subsets used in our paper. Each set targets medical vision–language QA and supports RLVR/SFT training.

Dataset	Generator	Verifier	Modality	Description	Download
MedSynVQA	GLM-4.5V 108B	Qwen2.5-VL 72B	Image–Text	Full training set for medical VQA (used for RLVR).	🤗 HF
MedSynVQA-10K	GLM-4.5V 108B	Qwen2.5-VL 72B	Image–Text	10K-sample training subset for RLVR.	🤗 HF
MedSynVQA-5K	GLM-4.5V 108B	Qwen2.5-VL 72B	Image–Text	5K-sample training subset for RLVR.	🤗 HF
MedSynVQA-2K	GLM-4.5V 108B	Qwen2.5-VL 72B	Image–Text	2K-sample training subset for RLVR.	🤗 HF
MedSynVQA-1K	GLM-4.5V 108B	Qwen2.5-VL 72B	Image–Text	1K-sample training subset for RLVR.	🤗 HF
MedSynVQA-5K-qwen-glm	Qwen2.5-VL 72B	GLM-4.5V 108B	Image–Text	5K subset for generator and verifier choice ablation (GLM→Qwen generator, Qwen→GLM verifier).	🤗 HF
MedSynVQA-5K-internvl-glm	InternVL-3.5 38B	GLM-4.5V 108B	Image–Text	5K subset for generator choice ablation (InternVL→GLM verifier).	🤗 HF
MedSynVQA-5K-glm-glm	GLM-4.5V 108B	GLM-4.5V 108B	Image–Text	5K subset for verifier choice ablation (Qwen→GLM verifier).	🤗 HF
MedSynVQA-5K-no-verify	GLM-4.5V 108B	N/A	Image–Text	5K subset for verifier necessity ablation (no verification step).	🤗 HF
MedSynVQA-5K-PMC-style	GLM-4.5V 108B	N/A	Image–Text	5K subset generated with PMC-VQA–style prompts.	🤗 HF
MedSynVQA-5K-SFT	GLM-4.5V 108B	N/A	Image–Text	5K subset generated for SFT training.	🤗 HF

Dataset Usage

from datasets import load_dataset

# Load evaluation dataset
eval_dataset = load_dataset("UCSC-VLAA/MedVLThinker-Eval")

# Load training dataset
train_dataset = load_dataset("MedVLSynther/MedSynVQA-13K")

Dataset details and preparation of your own

Data Format

All train datasets follow a unified format:

{
    "images": [PIL.Image],           # List of images                           
    "question": str,                 # Question text
    "options": Dict[str, str],       # Multiple choice options
    "answer_label": str,             # Correct answer label (A, B, C, D, E)
    "answer": str,                   # Full answer text
    "reasoning": str,                # Chain-of-thought reasoning (optional)
    "dataset_name": str,             # Source dataset name
    "dataset_index": int             # Unique sample identifier
}