Data card — HTR corpus and HTR-Benchmark (draft, 2026-11-09)
Summary
Paired data: source = a passage rewritten by a large language model (Claude Opus 5.5 primary; Sonnet 5.5 / Haiku 5.5
held-out), target = the original English passage from a document published before 2018-01-01 whose licence permits
the operations we perform. Both sides are stored in placeholder form (citations ⟦Cn⟧, math ⟦Mn⟧, links ⟦Un⟧; see
src/htr/parse/protect.py, protocol version recorded per pair). Metadata per pair: discipline, publication year/era,
section role, source archive, licence code, venue, author group, rewriting model, prompt version, protocol version.
Sources (all acquired through official channels, rate-limited, rights recorded per document)
| Source | Access route | Date evidence | Licence evidence | Scale observed |
|---|---|---|---|---|
| PMC Open Access Subset | AWS Open Data bucket pmc-oa-opendata (per-article JSON + JATS XML; MD5 verified) |
JATS <pub-date> (ppub → collection → epub) |
JSON license_code cross-checked with JATS <license xlink:href> |
pilot 1 180 docs; Medium ingestion ~10 000 |
| arXiv (q-fin, econ; others possible) | OAI-PMH arXivRaw metadata + e-print source of the last pre-2018 version (3 s between requests) |
v1 submission date | OAI license URI (only CC licences kept; the arXiv non-exclusive licence is excluded) |
162 eligible q-fin papers 2008–2017 |
| CC BY economics & finance journals (16 verified: MDPI Econometrics, JRFM, IJFS, Risks, Economies, Games; SpringerOpen Financial Innovation, J. Economic Structures, IZA journals, LAER, SJES, JIE) | Crossref REST (polite pool) + publisher PDF via Crossref text-mining links | Crossref published-print → published-online → issued |
Crossref license[] (content-version vor) |
1 730 eligible articles |
| Project Gutenberg (reference / scholarly non-fiction, LoCC Q R T H B G L Z) | catalog feed + official mirror gutenberg.pglaf.org |
front matter year (imprint / copyright / roman numerals); else author death year as an upper bound (flagged) | public domain in the US (catalog + licence header) | pool 7 799; 500 requested |
| NCBI Bookshelf open-access books | FTP pub/litarch (BITS XML) |
BITS <pub-date> |
BITS <permissions> (CC BY / public domain NIH works; UNKNOWN rejected) |
3 941 pre-2018 books; 300 requested |
| OpenStax / CNX pre-2018 editions | Internet Archive items; date and CC statement read inside the PDF | PDF copyright page | PDF licence statement | 5 trial items |
Excluded by policy: TDM author manuscripts, CC-BY-ND / CC-BY-NC-ND, UNKNOWN/OTHER-OA, the arXiv non-exclusive
licence, retracted articles, non-English text, near-duplicates (MinHash Jaccard ≥ 0.8), references/acknowledgements/
funding/ethics sections, captions and list items (kept in the parsed store, never eligible).
Processing
JATS/BITS, LaTeX, PDF and plain-text parsers → paragraphs with offsets into the normalised document text → eligibility
(60–400 words, ≥ 2 sentences, English ≥ 0.9, quality ≥ 0.6) → protected-span placeholders → LLM rewrite (variants A–E,
prompts rw-1.0) → automated validation (val-1.0: language, length ratio, near-copy, placeholder multiset, numbers,
entities, negation/hedging deltas, structure for D, formatting artefacts, MiniLM semantic similarity) → quarantine of
failures (kept, never deleted) → leakage-safe splits (document + near-duplicate cluster + author group; held-out discipline
and period; pilot assignments frozen).
Release policy (HTR-Benchmark)
- Public subset: pairs whose document licence is
CC0,PUBLIC-DOMAINorCC-BY— redistributed with attribution (title, authors, venue, DOI/PMCID/arXiv id/Gutenberg id, licence URI, "modified: segmented and placeholder-protected"). - Research subset (
CC-BY-SA,CC-BY-NC,CC-BY-NC-SA): metadata + identifiers + scripts to rebuild locally; text shared only under the matching licence terms (SA → CC BY-SA; NC → non-commercial research). - Never released: documents of tiers C/X, quarantined pairs' rewrites of non-releasable originals, provider request payloads beyond what the schema records.
- LLM rewrites are our outputs under Anthropic's commercial terms; their redistribution and use for training are recorded as a
compliance item to confirm before publication (
docs/RISK_REGISTER.md). - Takedown: a retraction or licence change at the source triggers removal from the next release and a dated note in
docs/DECISION_LOG.md.
Known biases and limitations
- Biomedicine dominates PMC; economics/finance, mathematics, physics and books are deliberate strata but smaller.
- Discipline labels are heuristic (journal title / subject categories / arXiv categories / LoCC classes).
- Pre-1950 PMC material is OCR and rarely passes quality; historical variation comes mostly from books, whose dates may be upper bounds.
- Validation heuristics: spelled-out numbers count as missing; entity matching is a proxy.
- The rewrites reflect a specific model generation (Claude 5.5 family, late 2026) and six prompt variants.
Citation and acknowledgement
Acknowledge NLM/NCBI (PMC Article Datasets on AWS, accessed 2026-11), arXiv, Crossref, the publishers of the CC BY
journals, Project Gutenberg, the Internet Archive and OpenStax (Rice University). See LICENSE_MATRIX.md.