The batch files that field pointed at are no longer shipped with the benchmark,
and every question still carries its generation batch in qa_source, so the field
was a dangling path. The card's field table follows.
Drop the last unrelated document and keep one directory per paper
Remove the NLP shared-task overview (2309.11285), the only corpus paper with
no LHC connection and no INSPIRE record, and remove the earlier version
directory of the duplicated paper so that each corpus paper has exactly one
directory. The corpus is now 2,771 papers (2,338 referenced by the metadata
files plus 433 further LHC documents) in 2,771 directories. Record every
removed paper with its reason in removed_papers.json, alongside the 57
INSPIRE records already moved to excluded_records.json.
Remove the astronomy name-collision papers and the unrelated documents
The corpus query matched the string "ATLAS" and therefore pulled in 58 papers
of the Herschel-ATLAS astronomy survey, and three further documents have no
connection to the LHC (a galaxy-ring paper, a black-hole X-ray binary and an
LLM technical report). Delete those 61 papers from arxiv/, rebuild
corpus_manifest.jsonl (2,772 rows), and move the 57 ATLAS records that
referred only to Herschel papers into excluded_records.json so the removal is
auditable rather than silent. The remaining corpus is 2,772 papers: 2,338
referenced by the metadata files and 434 further LHC documents. One unrelated
NLP shared-task overview (2309.11285) is left in place and flagged in the
card. Refreshed counts: ATLAS records 1,734 to 1,677.
Document the corpus manifest and the composition issues
Describe the manifest the release now ships: one row per corpus paper with its
INSPIRE title, collaborations, document type and year, plus flags for the
metadata snapshot and for INSPIRE coverage. Record what reconciling the
release showed: 58 Herschel-ATLAS astronomy papers that share the name, 58
sub-detector papers recorded under a group name, four documents unrelated to
the LHC, one paper stored twice, and the fact that the 96-paper indexed
subset used in the paper contains none of them. Correct the failure mode of
failed_*.csv: records with no arXiv source, not text-extraction failures.
Scope the license: our own artifacts under MIT, third-party paper texts under
their original terms.
Add the corpus manifest the dataset card describes
The card enumerates the corpus through corpus_manifest.jsonl - 2,833 rows, one
per paper, with INSPIRE title, collaborations, document type and year, plus flags
for whether the paper appears in ATLAS.json or CMS.json and whether an INSPIRE
record was found - but the file itself had not been committed.
Describe the corpus, the indexed subset and the benchmark separately
The card described a 96-paper corpus while arxiv/ holds 2,834 converted
papers, which contradicted both the repository contents and the paper. Split
the three roles and state the numbers as they reconcile: 3,646 INSPIRE
records, 3,250 with an arXiv id referring to 2,392 distinct papers (all
present), 396 records without an id listed with their DOIs, and 2,834
converted papers, i.e. the metadata papers plus 441 further ATLAS/CMS
documents. Correct the failure mode: the failed CSVs are records with no
arXiv source, not text-extraction failures. Scope the license: our own
artifacts under MIT, the third-party paper texts under their original terms.
Sync benchmark.jsonl with the benchmark repository
The generation-provenance paths in `qa_source_file` now name the current
locations under questions/ rather than the pre-restructure ones, and the two
spellings of the same batch are unified. No other field changes.
Rename the question set to benchmark.jsonl and add a dataset card
The benchmark repository now names its released question set benchmark.jsonl in
both benchmarks, so the dataset file matches. The card documents the columns,
the four genotypes (renumbered from the earlier seven-type taxonomy, with the
old-to-new mapping spelled out) and the retrievability gate. It also drops the
README.md line from .gitignore, which is what had kept this repository from
carrying a dataset card at all.
The four genotypes were numbered T3/T4/T5/T7 as leftovers of a seven-type
generation taxonomy. Number them sequentially to match the released paper:
T1 complementary bridge, T2 cross-document aggregation, T3 heterogeneous
multi-hop, T4 three-retriever. Only the `type` field changes on each of the
100 records.
Publish the prebuilt indexes
daweifufu
a33218e
The graph and vector indexes behind the reported scores, so the benchmark can be
run without paying for a full re-index. Entity and relationship extraction is an
LLM pass and does not reproduce bit for bit.
graph.pkl holds 24,909 nodes and 61,874 edges, which are the counts in the
paper's benchmark table; entities.pkl and relationships.pkl carry the FAISS
IndexFlatIP for those same nodes and edges, vectorstore/ carries the chunk index.
*.faiss needed an LFS rule (only *.pkl was covered), otherwise the 19.5 MB
index.faiss would have been committed as a plain blob.
LHCpublications: ATLAS+CMS corpus with the benchmark_100 QA set
