BiTaNan commited on
Commit
b67a8e7
Β·
verified Β·
1 Parent(s): e097eb3

Add dataset card and usage guide

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