Hugging Face's logo

Datasets:
Plans11
/
AncientScripts-Parallel-Corpus

Dataset card Files
xet
 Community
AncientScripts-Parallel-Corpus / BUILD_LOG.md
Plans11's picture
Plans11
AncientScripts Parallel Corpus v1.0.0: BUILD_LOG.md
953d8e3 verified
preview code
|
Raw History Blame Contribute Delete
6.82 kB

AncientScripts Parallel Corpus v1.0.0 — Build Log

Build date: 2026-10-19
Target: 25,000 rows
Final: 26,467 rows (unique, after SHA-256 dedup; 2 duplicates removed from 26,469 loaded)

SHA-256: 89b6bd5576636122630cf611694b72055a7c9c1d15cfbc81399b39366e97eea2

Final Counts by Writing System

Writing System Rows Sources
Latin 13,614 Pilot Caesar (404) + Latin extension (13,210)
Greek (polytonic) 6,917 Odyssey/Plato (3,938) + Iliad line-level (2,979)
Cuneiform 4,341 RIAo (4,343 loaded, 2 dups removed)
Chinese (Traditional) 894 Pilot TTC (81) + Analects (498) + Mencius/etc (315)
Devanagari (Sanskrit) 701 Bhagavad-Gita
Total 26,467

Build Order & Yields

1. Latin — 13,210 rows

Builder: build_latin.py (PerseusDL/canonical-latinLit pipeline)
Sources: Latin TEI (PerseusDL, CC BY-SA); English translations PD-verified per-text (pre-1930 imprint dates from TEI headers).
Coverage: Plautus (20 plays), Terence (6 plays), Caesar (Civil Wars), Catullus (Carmina), Cicero (speeches against Verres, etc.).
Decisions:

  • Excluded Caesar's Gallic War (already in pilot; avoids duplication).
  • Denylist for non-PD/unverifiable translations (Way 1955, Spencer 1935-38, Mahoney modern, etc.).
  • Guards: no zero-key merges, length-ratio ≤4.0, max 12k chars, 415 documented drops.
  • transliteration=null (alphabetic, documented).

2. Greek — 3,938 rows (Odyssey/Plato) + 2,979 rows (Iliad) = 6,917 total

Builder: child subagent (Odyssey/Plato); build_iliad2.py (Iliad, by coordinator)
Sources: Open Greek and Latin TEI (CC BY-SA 4.0); English translations PD-verified individually (pre-1929 imprints from __cts__.xml).
Coverage:

  • Homer Odyssey (24 books, 2,431 rows): Murray OCT 1919 Greek + Murray Loeb 1919 English, ~5-line chunks via shared milestones.
  • Homer Iliad (24 books, 24 rows book-level + 2,979 rows line-level): Monro & Allen OCT Greek + Butler 1898 English (book-level); Murray Loeb 1924-25 English (line-level, PD by age, discovered in OGL as perseus-eng3).
  • Plato (35 works, 1,473 rows): Burnet OCT Greek + Loeb English (Fowler 1914-26, Lamb 1924-27, Bury 1926-29), Stephanus-page level.
  • Plato Republic (10 books, 10 rows): Burnet OCT + Jowett 1892, book-level. Decisions:
  • Excluded Shorey Loeb Republic (1935-37, not PD).
  • Iliad line-level uses Murray 1924-25 (not Butler) for 5-line alignment via English milestones to Greek <l n="X"> line numbers.
  • transliteration=null (alphabetic, documented).

3. Classical Chinese — 498 rows (Analects) + 315 rows (Mencius/etc) = 813 total (+81 pilot = 894)

Builder: build_chinese.py (Analects); child subagent for extension
Sources: zh.wikisource.org (CC BY-SA transcription of PD text); James Legge English (1861, PD via Gutenberg).
Coverage:

  • Analects (論語) 20 books, 498 chapters, chapter-level alignment.
  • Mencius (孟子) 260 chapters + Great Learning (大學) 16 paragraphs + Doctrine of the Mean (中庸) 39 paragraphs = 315 rows. Decisions:
  • Excluded non-canonical 論語/知道第二十二 page.
  • transliteration=pinyin (machine-generated, flagged in notes).
  • ctext.org FORBIDDEN (not used).

4. Sanskrit — 701 rows (Bhagavad-Gita)

Builder: child subagent
Sources: sa.wikisource.org transcription of Annie Besant & Bhagavan Das, The Bhagavad-Gita (1905, PD).
Coverage: All 18 chapters, 701 verses, verse-level alignment.
Decisions:

  • Rejected Telang SBE 1882 (no verse numbers) and Arnold Song Celestial (no chapters).
  • Speaker attributions stripped from both sides.
  • transliteration=null? [VERIFY — Devanagari is alphasyllabary, may need IAST]

5. Cuneiform (Akkadian) — 4,341 rows (4,343 loaded, 2 dups removed)

Builder: build_cuneiform/ (fetch_riao_tei.py, parse_riao_tei.py, atf2unicode.py, build_cuneiform.py)
Sources: RIAo (Royal Inscriptions of Assyria online) via LMU Munich ORACC mirror (1,927 TEI files fetched, 0 failed). CC BY-SA 3.0 (project page) / CC0 (bulk metadata); using CC BY-SA 3.0.
Coverage: 1,458 texts yielded 4,343 translation segments; 469 texts had no English translations (skipped).
Method:

  • Each row = one RIAo translation segment, aligned via explicit xtr:sref/xtr:eref.
  • ATF→Unicode via hybrid mapper: MIT mapping.tsv (1123 entries) + Unicode names (1033 signs) + curated aliases + determinative table.
  • Coverage: ~97% of sign tokens mapped; '?' marks unmapped (loud, documented); '▯' marks broken signs.
  • transliteration=ATF (not null; cuneiform needs it). Decisions:
  • RIAo chosen over BALT (BALT has no English translations).
  • ETCSL translations FORBIDDEN (not used).
  • Dropped segments with no real script content (all broken).

6. Egyptian — SKIPPED

Reason: No bulk source with hieroglyphic + transliteration + English aligned. TLA requires registration; Gutenberg has English only; PDFs are not machine-readable. Documented honestly.

Hard Constraints Compliance

  • Public-domain or academic-open ONLY. No copyrighted modern translations.
  • ctext.org FORBIDDEN — not used.
  • ETCSL translations FORBIDDEN — not used.
  • CC BY-SA sources attributed in translation_source field.
  • SHA-256 dedup on (script_text, translation).
  • Seed-1147 shuffle.
  • manifest.json with per-source counts, SHA-256, license summary.
  • No padding with junk. Actuals documented honestly.

Problems Encountered

  1. Greek TEI milestone mismatch (Iliad): The Iliad Greek has sparse milestones (every ~30 lines) while Murray English has dense milestones (every 5 lines); only 5 coincide. Solved by extracting Greek <l n="X"> line elements (15,690 lines) and aligning to English 5-line milestone intervals.

  2. Cuneiform ATF→Unicode coverage: ~3% of sign tokens unmapped (rare signs like GUN, MURUB₄, indexed variants BAD₅/MIN₃). Marked as '?' (loud, documented) rather than guessed. The transliteration field preserves the full ATF, so no information is lost.

  3. RIAo fetch reliability: The LMU Munich mirror was slow (~20 files/min). All 1,927 TEI files fetched with 0 failures using polite 0.15s delays.

  4. Sanskrit verse alignment: Solved by using Besant & Das 1905 (verse-numbered) instead of Telang (no verses) or Arnold (no chapters).

Files Delivered

  • ancient-scripts-25k.jsonl (26,467 rows, SHA-256: 89b6bd5576636122630cf611694b72055a7c9c1d15cfbc81399b39366e97eea2)
  • manifest.json (per-source counts, SHA-256, license summary)
  • BUILD_LOG.md (this file)
  • parts/ (individual part files + NOTES.md for each)
  • build_cuneiform/ (reusable ATF→Unicode mapper + TEI parser)
  • build_iliad2.py, build_chinese.py, assemble_corpus.py (reusable builders)