Release metadata-only PosterBench dataset
Browse files- README.md +81 -0
- data/test.jsonl +0 -0
README.md
ADDED
|
@@ -0,0 +1,81 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: mit
|
| 3 |
+
pretty_name: PosterBench
|
| 4 |
+
language:
|
| 5 |
+
- en
|
| 6 |
+
tags:
|
| 7 |
+
- benchmark
|
| 8 |
+
- academic-posters
|
| 9 |
+
- poster-generation
|
| 10 |
+
- metadata-only
|
| 11 |
+
- arxiv:2608.13560
|
| 12 |
+
size_categories:
|
| 13 |
+
- n<1K
|
| 14 |
+
configs:
|
| 15 |
+
- config_name: default
|
| 16 |
+
data_files:
|
| 17 |
+
- split: test
|
| 18 |
+
path: data/test.jsonl
|
| 19 |
+
---
|
| 20 |
+
|
| 21 |
+
# PosterBench
|
| 22 |
+
|
| 23 |
+
PosterBench is a 100-paper, image-native benchmark for academic poster
|
| 24 |
+
generation introduced in [AutoDesign: Meta-Harness Optimization for
|
| 25 |
+
Long-Horizon Agentic Design](https://arxiv.org/abs/2608.13560).
|
| 26 |
+
|
| 27 |
+
This Hugging Face release is **metadata only**. It does not host or redistribute
|
| 28 |
+
the underlying paper PDFs, full text, abstracts, figures, tables, or
|
| 29 |
+
thumbnails. Each row identifies the exact benchmark paper version and records
|
| 30 |
+
the official landing page, access policy, license information where available,
|
| 31 |
+
and expected SHA-256.
|
| 32 |
+
|
| 33 |
+
## Load
|
| 34 |
+
|
| 35 |
+
```python
|
| 36 |
+
from datasets import load_dataset
|
| 37 |
+
|
| 38 |
+
dataset = load_dataset("YaxinLuo/PosterBench", split="test")
|
| 39 |
+
```
|
| 40 |
+
|
| 41 |
+
For the fixed 10-paper development subset, use
|
| 42 |
+
[`YaxinLuo/PosterBench-mini`](https://huggingface.co/datasets/YaxinLuo/PosterBench-mini).
|
| 43 |
+
|
| 44 |
+
## Dataset structure
|
| 45 |
+
|
| 46 |
+
The dataset contains 100 rows: 20 papers from each of five disciplines.
|
| 47 |
+
|
| 48 |
+
- `id`: stable `discipline/case_id` identifier
|
| 49 |
+
- `discipline`: benchmark discipline
|
| 50 |
+
- `case_id`: stable case identifier
|
| 51 |
+
- `title`, `authors`, `year`: bibliographic metadata
|
| 52 |
+
- `identifiers`: DOI, arXiv, and OpenAlex identifiers where available
|
| 53 |
+
- `landing_url`: official landing page
|
| 54 |
+
- `download_policy`: `open_access` or `manual_access_required`
|
| 55 |
+
- `pdf_url`: present only when the exact version has an approved Creative
|
| 56 |
+
Commons or public-domain license
|
| 57 |
+
- `license_url`: license for that exact paper version, when verified
|
| 58 |
+
- `expected_sha256`: SHA-256 of the exact version used by the benchmark
|
| 59 |
+
- `schema_version`: manifest schema version
|
| 60 |
+
|
| 61 |
+
The benchmark downloader automatically downloads only records with an HTTPS
|
| 62 |
+
PDF URL, an approved Creative Commons or public-domain license, and an exact
|
| 63 |
+
hash match. All other records require users to obtain the paper independently
|
| 64 |
+
through a lawful source available to them.
|
| 65 |
+
|
| 66 |
+
## License and rights boundary
|
| 67 |
+
|
| 68 |
+
The MIT license for this repository applies only to the PosterBench metadata,
|
| 69 |
+
benchmark-specific annotations, and documentation authored by the AutoDesign
|
| 70 |
+
project. It does **not** apply to or grant any rights in the underlying papers
|
| 71 |
+
or their contents. Each paper remains subject to its own copyright and license.
|
| 72 |
+
Users are responsible for confirming that their access and use are lawful.
|
| 73 |
+
|
| 74 |
+
## Source and reproducibility
|
| 75 |
+
|
| 76 |
+
- [AutoDesign repository](https://github.com/Yaxin9Luo/AutoDesign)
|
| 77 |
+
- [PosterBench evaluation guide](https://github.com/Yaxin9Luo/AutoDesign/tree/main/eval)
|
| 78 |
+
- [Canonical manifest](https://github.com/Yaxin9Luo/AutoDesign/blob/main/eval/benchmark_manifest.jsonl)
|
| 79 |
+
- [Paper page](https://huggingface.co/papers/2608.13560)
|
| 80 |
+
|
| 81 |
+
Please cite the AutoDesign paper when using PosterBench.
|
data/test.jsonl
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|