LHCpublications

ATLAS + CMS publications corpus (INSPIRE metadata + PDFs) with OCR-extracted markdown and figures, plus a 100-question multi-document benchmark for retrieval / agentic RAG.

简体中文版本 / Simplified Chinese version

Contents

Path	Description
arxiv/<id>/<id>.pdf	PDF snapshot (Git LFS)
arxiv/<id>/<id>.md	OCR-extracted markdown
arxiv/<id>/*.jpg	Figures extracted during OCR
ATLAS.json, CMS.json	INSPIRE metadata records for the ATLAS / CMS corpora
failed_ATLAS.csv, failed_CMS.csv	Records whose PDF or metadata fetch failed
benchmark_100.jsonl	100-question benchmark (one JSON object per line)

Corpus size: 2834 papers. All questions and evidence are in English.

Benchmark: benchmark_100.jsonl

100 questions, balanced across four multi-document types:

Type	n	Meaning
T3	25	Complementary bridging — two or three papers each hold a disjoint part of the answer; must be combined
T4	25	Cross-document aggregation/comparison — the same observable collected across ≥2 papers, then compared
T5	25	Heterogeneous multi-hop — different hops require different retrieval modalities
T7	25	Three-modality — numeric value + entity relation + cross-document aggregation, all required

Every question requires evidence from more than one paper and is grounded in evidence snippets that were verified retrievable from the corpus.

Record fields

Field	Description
question	Question text given to the system under test
type	Question type (see table above)
arxiv_ids	Papers needed to answer
arxiv_id	Primary paper
gold_answer / reference_answer	Answer key (identical in this release)
evidence	Per-hop grounding: list of {arxiv_id, section, snippet}
cluster	Source paper cluster the question was generated from
qa_source / qa_source_file	Provenance: which generation batch the question came from (present on all 100 records)
benchmark / benchmark_role	Dataset name (benchmark_100) / role (main)
gate_pass	Whether the question passed the retrievability gate

Key order and field set are identical across all 100 records for diffability.

Citation

If you use this corpus or the benchmark, please cite:

@misc{lhcpublications2026,
  title        = {LHCpublications: An ATLAS and CMS Publication Corpus with a 100-Question Multi-Document Benchmark},
  author       = {Fu, Dawei and Jiang, Cheng},
  year         = {2026},
  howpublished = {Hugging Face Datasets},
  url          = {https://huggingface.co/datasets/PKUfudawei/LHCpublications}
}

The underlying publications are the property of the ATLAS and CMS collaborations; please also respect the licensing of the original papers when redistributing text or figures extracted from them.