Gated dataset. Approve to view/download. License: public domain across the whole repo (Corpus Reformatorum editions, archive.org/HathiTrust public-domain-curated scans, pre-1929 public-domain imprints/critical editions); the OCR/repair/lemmata/apparatus derivations are released under the same terms.

Reformation (REF) — reception corpus, page-grain OCR

Reformation-era reception corpus (Calvin, Luther, Melanchthon, Bellarmine, Erasmus, Bucer, Bullinger; c. 1517-1564 CE and their early printed/critical editions). Tesseract v4 page-grain OCR at 200 DPI, plus derived apparatus, lemmata, content-type classification, clean-text projections, and per-author repair sidecars (dehyphenation, footnote-apparatus repair, long-s normalization). All REF interconnects run temporal_direction=reverse (reception of Scripture, not influence on it).

Shared cross-author tables

config	rows	what
`ref_page_coords`	101,628	page-grain coordinate/bbox registry
`ref_works`	1,589	work registry, 177 volumes across 8 lane-added authors
`ref_segments`	870,908	canonical segment table (T0 basis for embeddings)
`ref_apparatus`	253,676	decomposed critical apparatus
`ref_general_indices`	42,066	index/navigation surface
`ref_pervolume`	104,851	per-volume rollup
`ref_residual_closure`	1,229	frozen-vintage residual closure sidecar (source JSONL deleted in a later reclaim wave; parquet is the artifact of record)
`ref_seg_binding`	766,848	segment-binding table
`ref_segment_embeddings`	766,848	LaBSE segment embeddings (PD Reformation text; not contaminated by catlib k-lineage)
`ref_segment_embeddings_v2_norm_text`	870,908	v2 normalized-text embedding variant
`ref_claim_embeddings`	67,992	claims-plane embeddings (claims cascade PARKED; table published, not tool-registered)
`ref_claim_relations`	6,440	claim-relation edges (PARKED, excluded this wave)
`ref_controversy_edges`	727	controversy-edge table (PARKED, excluded this wave)
`ref_citations_nonscripture_tiered`	1,826	A1-REF/A2-REF evidence-tiered non-scripture citation resolution v1d (tool spec `rc_ref_citations_nonscripture` authored, registration pending)
`zero_label_refs`	62,460	zero-label diagnostic sidecar
`ref_index_kind`	211,461	I2-REF pre-joined typed printed-index/nav-TOC layer over 5 source configs (published wave 2, `rc_ref_index_lookup`/`rc_ref_nav_toc`)
`tll_bridge_ref`	3,484,251	T12c-REF author x lemma_norm lexicon bridge (published wave 2; bridge layer, not a tenant tool this wave)

Per-author tables

Each of the 7 authors carries the same 8-config pattern: base registry, OCR, clean-text projection, content-type classification, lemmata, repair sidecar, repair-sidecar footnotes, repair-sidecar long-s normalization. Calvin additionally carries two single-file supplement configs frozen from an earlier build (calvin_apparatus_supplement, calvin_nav_supplement).

author	base	ocr	clean_text	content_type	lemmata	repair_sidecar	repair_sidecar_footnotes	repair_sidecar_long_s
calvin	23,460	315,603	315,603	19,489	225,108	115,277	3,694	4,588
luther	54,166	940,629	940,629	42,942	613,191	210,881	3,991	9,597
melanchthon	15,296	324,481	324,481	16,191	236,317	108,369	4,325	1,415
bellarmine	7,852	128,973	128,973	2,501	99,936	57,382	46	47,076
erasmus	12,201	281,115	281,115	8,069	242,163	125,803	867	97,294
bucer	4,097	41,745	41,745	473	34,746	19,104	10	13,799
bullinger	9,721	95,713	95,713	1,897	80,445	43,043	22	31,100

Calvin supplements: calvin_apparatus_supplement = 10,642 rows; calvin_nav_supplement = 855 rows.

Row counts above are the ontology/manifest/assets.yaml SOT pins (hf.lock-consistent). Five configs were spot-verified this leg by downloading each config's parquet via hf download (narrow --include pull, never anonymous/httpfs) into the session scratchpad and counting rows with local duckdb, then deleting the download: ref_works (1,589), bucer_content_type (473), bullinger_repair_sidecar_footnotes (22), calvin_nav_supplement (855), ref_controversy_edges (727) — all matched the pinned count exactly. No token touched a URL; nothing was printed.

feats_grammar_bridge (T12b-REF-BUILD, 2026-09-12): token-grain Stanza FEATS → grammar-construction bridge for the REF rail, 158 volumes / all 7 REF authors, 40,227,228 rows, 24 columns (same schema as unified-grammar/patristics-scholastics feats_grammar_bridge). Partial build: 91 worklist volumes and the German-language increment (featslemma-refde-01..04, __deu__ page-range splits of CR_22/CR_23/CR_26/WA_34_ii) remain outstanding for a follow-up increment; CR_22/CR_23/CR_26 carry the corrected Latin page-range splits only. coverage_summary.json beside the parquet lists per-volume token sums (row count exact-matches their total). k=computed; facts only, no source prose. Receipt: W53_parity/lexicon/t12b_ref/T12B_REF_BUILD_RECEIPT_v1.md.

Method

Tesseract v4 page-grain OCR at 200 DPI. Per-author repair sidecars apply dehyphenation and a footnote-apparatus pass; repair_sidecar_long_s normalizes historical long-s (ſ) forms. clean_text is a per-volume clean-text projection of the corresponding _ocr config — the source text in _ocr is never overwritten (granularity preserved).

License

Public domain (k=public-domain) across the whole repo:

Calvin and Melanchthon: Corpus Reformatorum, 19th-century edition.
Luther: public-domain-curated archive.org / HathiTrust scans (workspace_pd_safe flag at acquisition).
Bellarmine, Erasmus, Bucer, Bullinger: pre-1929 public-domain imprints/critical editions.

Corrected 2026-07-31: the Weimarer Ausgabe (WA) volumes 61-72 register apparatus (Böhlau 1983-1990 edition, in copyright) is not in this repo — it was relocated to the private NuBerea/quarantine-works (wa_registers config). The Luther volumes retained here are the PD text only.

Two supplement configs (calvin_apparatus_supplement, ref_residual_closure) carry a frozen-vintage build exception (their source JSONL was deleted in a later reclaim wave; the published parquet is the frozen artifact of record) — this affects rebuildability, not licensing; both remain k=public-domain.

Known card gap (disclosed, not silently carried forward)

The live repo carries 75 data/<config> directories; the prior revision of this card (stub, pre-T2-REF) declared only ~18 config groups. The T2-REF Step 1 revision (2026-09-11) declared 73 live configs, verified against hf datasets info NuBerea/reformation (856 siblings, 73 distinct data/ top-level directories at that time) — the previously-undeclared set included all 21 T8b-REF per-author content_type/lemmata/repair_sidecar configs, the 7 REF-R1 <author>_repair_sidecar_long_s clean-text configs, the 7 <author>_clean_text configs, and ref_citations_nonscripture_tiered.

T2-REF Step 2 refresh (2026-09-12): re-verified against the live repo after PUBLISH-W53-WAVE2 (FABLE_GATE_PUBLISH-W53-WAVE2.json, ACCEPT) — hf datasets info NuBerea/reformation --format json now shows 858 siblings, 75 distinct data/ top-level directories (sha 3720d7c2d0dc645981f63d060ed10b0ab55c869d, private:true, gated:false, pre-flip). The 2 newly-live configs added this revision: ref_index_kind (211,461 rows, published 1592563a) and tll_bridge_ref (3,484,251 rows, published 1907c31c) — both confirmed present in data/ and matching the assets.yaml SOT row-count pins exactly. The 7 <author>_clean_text configs were republished (second-pass) in the same wave with unchanged row counts (verified against assets.yaml, no table update needed for those rows). No SPEC.md or raw parquet is referenced or linked from this card, per the gated regime.

Planned (not yet published)

zwingli_clean_text (REF-ZWINGLI-PROJECTION, 2026-09-12): Zwingli has no OCR volume in the estate, so it does not carry the 8-config OCR pattern of the other 7 authors — only ref_works/ref_segments/ref_apparatus (0-row placeholder) via the lane-2 UZH born-digital build. zwingli_clean_text is a segment-grain clean-text projection over the live ref_segments zwingli slice (8,182 rows), stripping only the UZH page template's own navigation chrome (prev/index/next header, paginated jump-to-page footer) — never the edition's own citations or letter datelines. k=public-domain (Corpus Reformatorum 88-96, 1905-1968 imprints, PD by age; UZH site copyright notice is chrome, not a claim over the edition text — owner ruling 2026-09-12). Staged locally (huggingface/candidates/build/reformation/data/zwingli_clean_text/zwingli.parquet, 8,182 rows); not yet in configs: above or on HF — added to both on publish. See ~/Documents/OCR/ecrp_v1/W53_parity/reformers/ZWINGLI_PROJECTION_RECEIPT_v1.md.

Gated dataset. Approve to view/download. License: public domain.

Downloads last month: 395

Total file size:

10.9 GB