Datasets:
Modalities:
Image
Languages:
English
Size:
10K<n<100K
Tags:
medical
medical-vlm
vision-language
adversarial-benchmark
evidence-grounded-reasoning
clinical-report-generation
License:
File size: 7,153 Bytes
b67a8e7 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 | ---
pretty_name: Med-R2 Bench
language:
- en
license: other
size_categories:
- 10K<n<100K
task_categories:
- visual-question-answering
- image-to-text
- question-answering
tags:
- medical
- medical-vlm
- vision-language
- adversarial-benchmark
- evidence-grounded-reasoning
- clinical-report-generation
configs:
- config_name: default
data_files:
- split: train
path: train/*.jsonl
- split: test
path: test/*.jsonl
---
# π©Ί Med-R2 Bench
### An adversarial benchmark for evidence-grounded reasoning in medical vision-language models
Med-R2 Bench evaluates whether a medical VLM can make decisions from visual evidence instead of relying on superficial language patterns. The benchmark combines image-quality perturbations, anatomy-level reasoning, lesion-level reasoning, clinical-report generation, and multi-step joint QA in one consistent release.
π **Paper:** *Med-R2: An Adversarial Benchmark for Evidence-Grounded Reasoning in Medical VLMs* (EMNLP 2026 Main Conference)
π **Project repository:** [BiTaNannnnn/Med-R2-Bench on GitHub](https://github.com/BiTaNannnnn/Med-R2-Bench)
---
## β¨ What is in the dataset?
The public release contains **99,997 English JSONL QA examples** and the image files referenced by those examples.
| Split | Level 1 | Level 3 | Level 4 | Level 5 | Joint QA | Total |
| --- | ---: | ---: | ---: | ---: | ---: | ---: |
| π§ͺ Test | 2,640 | 7,316 | 11,140 | 600 | 500 | **22,196** |
| ποΈ Train | 5,184 | 14,152 | 25,557 | 30,320 | 2,588 | **77,801** |
| π Total | **7,824** | **21,468** | **36,697** | **30,920** | **3,088** | **99,997** |
> Level 2 is not part of this benchmark release. The level numbering follows the original Med-R2 task design.
## π§ Evaluation tracks
| Track | Medical reasoning focus | Representative tasks |
| --- | --- | --- |
| **Level 1 Β· Image quality** | Detect whether an image has been corrupted or altered | Noise recognition, artifact recognition |
| **Level 3 Β· Anatomy** | Identify organs and reason about location or erased evidence | Organ recognition, organ location, organ eraser identification |
| **Level 4 Β· Lesions** | Reason about body parts, lesions, boxes, and lesion size | Body-part recognition, lesion recognition, multi-box reasoning, lesion size |
| **Level 5 Β· Clinical reports** | Generate a report under localized or image-level interference | Box report, noise interference, artifact interference |
| **Joint QA Β· Multi-step reasoning** | Answer a sequence from modality to anatomy, abnormality, and diagnosis | Four-step sequential visual reasoning |
## βοΈ Adversarial diversity
Med-R2 is designed to test robustness under controlled visual changes. Depending on the task, examples include:
- β
**Positive cases:** the visual evidence and the question are consistent.
- β **Neutral / normal cases:** the image provides no abnormal evidence for the queried condition.
- β **Negative / adversarial cases:** the question, region, or visual evidence is intentionally mismatched.
- π₯ **Box interventions:** correct boxes, wrong boxes, or localized regions that redirect attention.
- π§Ή **Erased-region cases:** relevant anatomy or evidence is removed or occluded.
- π **Noise and artifact interventions:** image-level corruption is added while the underlying medical content is controlled.
This diversity makes it possible to measure not only average accuracy, but also whether a model remains evidence-grounded across positive, normal, and adversarial conditions.
## π Repository layout
```text
Med-R2-Bench/
βββ train/
β βββ visual_evidence_Level1.jsonl
β βββ visual_evidence_Level3.jsonl
β βββ visual_evidence_Level4.jsonl
β βββ visual_evidence_Level5.jsonl
β βββ visual_evidence_joint_qa.jsonl
βββ test/
β βββ same five JSONL files as train/
βββ Med-R2-Bench/
βββ dataset/ # Referenced generated and processed images
βββ images/ # Referenced visual-evidence and joint-QA images
```
The JSONL files are the benchmark annotations. The image files are released separately in the nested `Med-R2-Bench/` directory so that the `image_paths` field remains reproducible after downloading the repository.
## π§Ύ Annotation format
Each JSONL line is one example. The exact fields depend on the evaluation track, but common fields include:
| Field | Description |
| --- | --- |
| `image_paths` | Repository-relative image path, such as `Med-R2-Bench/dataset/...png` |
| `question` | The visual question or report-generation instruction |
| `options` | Multiple-choice options when the task is closed-ended |
| `answer` | Gold answer for the example |
| `task` | Fine-grained task name, such as `noise_recognition` or `lesion_recognition` |
| `type` | Data condition or adversarial bucket |
| `modality` | Imaging modality, when available |
Level 5 report examples use an open-ended `answer` field rather than multiple-choice options. Joint QA examples include the sequential prompts and gold answers for each reasoning step.
## π Loading the annotations
```python
from datasets import load_dataset
dataset = load_dataset(
"BiTaNan/Med-R2-Bench",
data_files={
"train": "train/*.jsonl",
"test": "test/*.jsonl",
},
)
print(dataset)
print(dataset["test"][0])
```
To resolve an image locally after downloading the repository:
```python
from pathlib import Path
repo_root = Path("/path/to/Med-R2-Bench")
image_file = repo_root / example["image_paths"]
```
The image paths are intentionally relative and contain no private server prefixes. The public package does not include model weights or prediction results.
## π Evaluation code and prompts
The benchmark-construction code, metric scripts, comparison scripts, and prompt notes are maintained in the companion GitHub repository:
π [https://github.com/BiTaNannnnn/Med-R2-Bench](https://github.com/BiTaNannnnn/Med-R2-Bench)
The Hugging Face repository is the **data release**. The GitHub repository is the **code and documentation release**.
## π Release and usage notes
- The released annotations use English prompts and answers.
- Absolute server paths have been removed from the public annotations.
- Model weights, private experiment outputs, API keys, and training logs are not included.
- Please check the licenses and terms of the underlying medical image sources before redistributing the data or using it in a new collection.
- This benchmark is intended for research evaluation and does not provide medical advice or a clinical diagnostic service.
## π Citation
```bibtex
@inproceedings{medr2-2026,
title = {Med-R2: An Adversarial Benchmark for Evidence-Grounded Reasoning in Medical VLMs},
author = {Ma, Wen and collaborators},
booktitle = {Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing},
year = {2026}
}
```
β If you use Med-R2 Bench, please cite the paper and link back to both the dataset and the [GitHub repository](https://github.com/BiTaNannnnn/Med-R2-Bench).
|