pscience674/ds-7f93

PreScience dataset (anonymized release for double-blind review)

Two domains, each with a temporally split train/test corpus:

Domain	arXiv categories	Train targets (Aug 2024 – Jul 2025)	Test targets (Aug 2025 – Jul 2026)	Papers (train / test)
prescience-ai (PreScience-AI)	cs.CL, cs.LG, cs.AI, cs.CV, cs.IR, cs.NE	49,351	67,575	482,420 / 629,936
prescience-quantum (PreScience-Quantum)	quant-ph, cond-mat, physics.optics, physics.atom-ph	13,290	18,220	371,813 / 442,167

Each <domain>/<split>/ directory contains:

• all_papers.json.gz: {"data": [paper, ...]}. Every paper in the split: target papers ("target" in roles) plus companion papers (influential references and author publication histories). Target records include corpus_id, arxiv_id, date, categories, title, abstract, authors (disambiguated author_id, name, publication_history, and bibliometrics snapshotted at the paper's date), key_references (influential references), citation_trajectory (monthly cumulative citation counts after publication), and, for PreScience-AI, topic labels (topics).
• sd2publications.json.gz: {"data": {...}}, a map from disambiguated author ID to the corpus IDs of that author's publications.

Download the files (e.g. hf download pscience674/ds-7f93 --repo-type dataset --local-dir prescience), decompress with gunzip -r ., place the domain directories under data/ in the code release, and pass --data_dir data/prescience-ai or --data_dir data/prescience-quantum to any script.

GRIT, GTR, and SPECTER2 embeddings are not included because of their size; the code release includes the script that computes them.

License: ODC-BY 1.0.