File size: 2,055 Bytes
daa31b5
99a5502
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
daa31b5
 
99a5502
 
 
 
 
a7ceed1
 
 
99a5502
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
---
license: cc-by-nc-4.0
pretty_name: PaperGuard
task_categories:
- image-text-to-text
- text-generation
language:
- en
size_categories:
- 1K<n<10K
tags:
- peer-review
- adversarial-robustness
- prompt-injection
- multimodal
- benchmark
configs:
- config_name: default
  data_files:
  - split: test
    path: data/test-*
---

# PaperGuard

A benchmark of academic papers (text + key figures) for evaluating the robustness of
multimodal AI peer-review systems.

**Paper:** [arXiv:2606.12716](https://arxiv.org/abs/2606.12716) — *Does AI Reviewer See the Full Picture? Attacking and Defending Multimodal Peer Review* · **Project Pages:** [paper-guard.github.io](https://paper-guard.github.io/)


## Schema (single `test` split)

| column | type | description |
|---|---|---|
| `paper_id` | string | paper identifier (e.g. `502`, `ICLR2020_10`, `1-26`) |
| `source` | string | one of `iclr_2017`, `AgentReview`, `F1000` |
| `title` | string | paper title (may be null) |
| `abstract` | string | abstract text |
| `num_figures` | int | number of figures present (0–2) |
| `paper_json` | string | the full original parsed-paper JSON (`{"name", "metadata"}`), verbatim |
| `method_figure` | image | the method figure (`<id>-1.png`), or null |
| `result_figure` | image | the result figure (`<id>-2.png`), or null |

## Usage

```python
from datasets import load_dataset

ds = load_dataset("rellabear/PaperGuard", split="test")
row = ds[0]
fig = row["method_figure"]   # PIL.Image or None
```

## License & Attribution

Released for **non-commercial research** under **CC-BY-NC-4.0**. Please cite the original sources:

- **F1000** — via [NLPeer](https://github.com/UKPLab/nlpeer) (Dycke et al., ACL 2023); F1000Research content is CC-BY.
- **iclr_2017** — via [PeerRead](https://github.com/allenai/PeerRead) (Kang et al., NAACL 2018).
- **AgentReview** — via [AgentReview](https://github.com/Ahren09/AgentReview) (Jin et al., EMNLP 2024).

ICLR/OpenReview-derived content remains subject to [OpenReview's terms](https://openreview.net/legal/terms).