CoolFace
Datasetpublic

YaxinLuo/PosterBench

PosterBench PosterBench is a 100-paper, image-native benchmark for academic poster generation introduced in AutoDesign: Meta-Harness Optimization for Long-Horizon Agentic Design. This Hugging Face release is metadata only. It does not host or redistribute the underlying paper PDFs, full text, abstracts, figures, tables, or thumbnails. Each row identifies the exact benchmark paper version and records the official landing page, access policy, license information where available… See the full description on the dataset page: https://huggingface.co/datasets/YaxinLuo/PosterBench.

sourceHugging Facemitupdated 1mo agoView on Hugging Face
0likes41downloads
Dataset Card

PosterBench

PosterBench is a 100-paper, image-native benchmark for academic poster generation introduced in AutoDesign: Meta-Harness Optimization for Long-Horizon Agentic Design.

This Hugging Face release is metadata only. It does not host or redistribute the underlying paper PDFs, full text, abstracts, figures, tables, or thumbnails. Each row identifies the exact benchmark paper version and records the official landing page, access policy, license information where available, and expected SHA-256.

Load

python
from datasets import load_dataset

dataset = load_dataset("YaxinLuo/PosterBench", split="test")

For the fixed 10-paper development subset, use `YaxinLuo/PosterBench-mini`.

Dataset structure

The dataset contains 100 rows: 20 papers from each of five disciplines.

  • id: stable discipline/case_id identifier
  • discipline: benchmark discipline
  • case_id: stable case identifier
  • title, authors, year: bibliographic metadata
  • identifiers: DOI, arXiv, and OpenAlex identifiers where available
  • landing_url: official landing page
  • download_policy: open_access or manual_access_required
  • pdf_url: present only when the exact version has an approved Creative Commons or public-domain license
  • license_url: license for that exact paper version, when verified
  • expected_sha256: SHA-256 of the exact version used by the benchmark
  • schema_version: manifest schema version

The benchmark downloader automatically downloads only records with an HTTPS PDF URL, an approved Creative Commons or public-domain license, and an exact hash match. All other records require users to obtain the paper independently through a lawful source available to them.

License and rights boundary

The MIT license for this repository applies only to the PosterBench metadata, benchmark-specific annotations, and documentation authored by the AutoDesign project. It does not apply to or grant any rights in the underlying papers or their contents. Each paper remains subject to its own copyright and license. Users are responsible for confirming that their access and use are lawful.

Source and reproducibility

Please cite the AutoDesign paper when using PosterBench.