README.md

metadata

license: other
license_name: rap-benchmark-composite-license
license_link: LICENSE
pretty_name: RAP Benchmark Dataset

RAP Benchmark Dataset

This is the companion data release for Research Attention Prediction (RAP). RAP evaluates whether an agent can estimate the distribution of newly posted papers across eight fixed research directions, then forecast the distribution in the following six months.

The package contains the frozen benchmark interface and the metadata-only Search snapshot.

Snapshot

Item	Count
Fields	278
Dev episodes	335
Test episodes	1,055
Fixed Search shards	1,390
Expanding Search shards	1,390
Directions per field	8

Layout

Path	Use
inputs/	Safe model-facing Dev/Test episode inputs
targets/	Evaluator-only recent/future counts, shares, and diagnostics
evaluator/	Combined episodes for scoring and no-API baselines; do not send to agents
data/	Public runner scope and Search-manifest indexes
search_shards/	Fixed and Expanding metadata-only Search corpora
DATA_CARD.md	Task semantics, leakage boundary, and aggregation rules
ARXIV_METADATA_NOTICE.md	Provenance and redistribution boundary for arXiv metadata

Each row in inputs/ contains the field, cutoff date, history-window metadata, future-window metadata, and the eight direction definitions. It intentionally does not contain history counts, recent shares, future counts, or future shares. Those evaluator-only values live in targets/ and evaluator/.

Search snapshot

• search_shards/fixed/ enforces [T-6m, T).
• search_shards/expanding/ enforces <T.
• Each shard contains exactly id, date, title, abstract, and categories.
• Each manifest records the shard SHA-256, row count, and date-bound checks.

The shards contain descriptive arXiv metadata. See ARXIV_METADATA_NOTICE.md for provenance and source terms.

Integrity

Verify the complete snapshot from its root:

sha256sum -c SHA256SUMS

The evaluator contract and scoring commands are documented in the companion RAP evaluation-code repository.

License

This dataset repository is marked other because it combines two provenance classes. Original RAP-derived artifacts are released under CC BY 4.0. The metadata-only Search shards retain their arXiv provenance and source terms. Read LICENSE and ARXIV_METADATA_NOTICE.md before redistributing or adapting the snapshot.

Citation

Please cite the RAP paper:

@misc{wu2026rap,
  title  = {RAP: Research Attention Prediction Reveals Target-Conditioned Evidence Acquisition Biases},
  author = {Wu, Yingqian and Liang, Jingcong and Wang, Siyuan and Yin, Zhenfei
            and Torr, Philip and Yu, Junchi and Wei, Zhongyu},
  year   = {2026},
  url    = {https://arxiv.org/abs/2609.10092}
}