Hugging Face's logo

Datasets:
Plans11
/
AncientScripts-Parallel-Corpus

Dataset card Files
xet
 Community
AncientScripts-Parallel-Corpus / SOURCES.md
Plans11's picture
Plans11
AncientScripts Parallel Corpus v1.0.0: SOURCES.md
b84fad5 verified
preview code
|
Raw History Blame Contribute Delete
8.91 kB

AncientScripts Parallel Corpus — Source Verdicts

Research date: 2026-10-19. Standard: ONLY public-domain or explicitly academic-open sources. No scraping of copyrighted modern translations. Every verdict below cites the source's own terms.

1. Sumerian (cuneiform) — CAUTION / GO via ORACC

ETCSL (Electronic Text Corpus of Sumerian Literature, Oxford): NO-GO without permission. Terms (http://etcsl.orinst.ox.ac.uk/edition2/usingbrowse.php): "There is no restriction on printing the compositions one-by-one for personal and educational use. However, permission must be obtained before including the compositions in a book, compendium, or similar for distribution or sale." A training dataset released on the Hub is a compendium for distribution → permission required. Do not ingest ETCSL translations.

ORACC (Open Richly Annotated Cuneiform Corpus): GO. License (https://oracc.museum.upenn.edu/doc/about/licensing/index.html): "released under a Creative Commons Attribution Share-Alike license… you are free to copy, distribute, transmit and adapt our work without permission" (attribute + share-alike). Bulk access: JSON API per project — http://oracc.org/[PROJECT]/manifest.json, corpus.json, corpusjson/[ID].json (documented at oracc.ub.uni-muenchen.de/doc/opendata/json). Formats include atf (transliteration), lem (lemmatized), and tr-en (English translation). 146 public projects. Yield: tens of thousands of texts with translations (varies by project; e.g. royal-inscription projects are translation-rich). Caveat: oracc.org and oracc.museum.upenn.edu were unreachable from our sandbox on 2026-10-19 (connection drops); the Munich mirror serves the project list but not all per-project JSON. Retry from another network or use per-project Zenodo dumps (e.g. BALT: 2,990 Babylonian texts, zenodo.org/records/17792290). Unicode: corpora store ATF transliteration (sign names like ŠUKU, BAN₂), NOT Unicode cuneiform. True Unicode cuneiform (U+12000 block, well supported in fonts) requires a sign-list mapping step — flagged as v2 work.

CDLI: GO (with limits). "Fully committed to open access… data… for download in many formats, and through the API… reused by researchers, but also for commercial endeavours" (cdli.earth). 300k+ artifacts. LIMIT: mostly transliteration + metadata; English translations are sparse. Best as a transliteration/metadata source, not for triples.

2. Akkadian (cuneiform) — GO via ORACC

Same verdict as Sumerian: ORACC is the primary source (CC BY-SA, JSON API, tr-en formats). Akkadian royal inscriptions (RINAP etc.) and the BALT Zenodo dump are translation-rich. CDLI as supplement. Same Unicode-cuneiform mapping caveat as above.

3. Egyptian hieroglyphs — CAUTION (needs more research)

Thesaurus Linguae Aegyptiae (TLA, Berlin-Brandenburg Academy): CAUTION. Requires free registration; the interface is search-based with no documented bulk API; and translations are in GERMAN, not English ("Texts are presented in transliteration and in German translation"). The TLA lemma list is CC BY-SA 4.0 (per hieroglyphica.org legal page), but full-text bulk extraction is not clearly permitted or technically available. Not suitable as-is for an English-parallel corpus. Unicode hieroglyphs (U+13000–U+1342F) have good font support — the blocker is licensing and access, not encoding. Alternative to investigate: the St. Andrews Corpus (transliteration + English translation + some hieroglyphic text) — license not yet verified. Pyramid Texts (Sethe, Univ. of Chicago) — likely public domain given age, not yet verified.

4. Sanskrit (Devanagari) — GO

Sacred Books of the East (1879–1910, ed. Max Müller): GO. Unambiguously public domain by age. LIMIT: English translations only — no Devanagari source text. Wikisource (sa.wikisource.org): GO with CC BY-SA 4.0 attribution on transcriptions (the underlying texts are PD). Verified live 2026-10-19: the Bhagavad Gita in Devanagari across 18 chapter subpages, verses numbered ॥१-N॥ in <poem> blocks. Unicode Devanagari: excellent support. Caveat found in pilot research: Edwin Arnold's "Song Celestial" (Gutenberg #2388, PD) has NO chapter divisions after CHAPTER I — it is one continuous poem, so it cannot cleanly align to verse- or chapter-numbered Devanagari. For verse-level Gita triples, a verse-numbered PD translation is still needed (Besant/Das 1905 or Telang SBE Vol. 8 — not yet pulled). IAST transliteration of Devanagari is deterministic and safe to generate.

5. Classical Chinese — GO via Wikisource; ctext.org is NO-GO

ctext.org (Chinese Text Project): NO-GO for bulk. Terms on every page: "Please note that the use of automatic download software on this site is strictly prohibited, and that users of such software are automatically banned without warning to save bandwidth." Explicit anti-scraping + site-content copyright claim. Do not use. Wikisource (zh.wikisource.org): GO (CC BY-SA transcription of PD text). Verified 2026-10-19: 道德經 (王弼本), all 81 chapters cleanly marked, Wang Bi commentary in {{*|…}} templates (strippable). Unicode Traditional Chinese: excellent. Pilot pairing: Legge 1891 translation (Project Gutenberg #216, PD — "You may copy it, give it away or re-use it under the terms of the Project Gutenberg License"). Chapter-level 1:1 alignment verified across all 81 chapters.

6. Ancient Greek — GO (script); translations need per-text check

Open Greek and Latin: GO. "Perseus has published 41 million words of Greek and 17 million words of Latin in curated TEI XML… available under a Creative Commons (CC-BY-SA) license. These texts were developed specifically so that third parties could download, modify and reuse them." GitHub: PerseusDL/canonical-greekLit, OpenGreekAndLatin/First1KGreek (+ Zenodo DOIs). Bulk: git clone. Unicode polytonic Greek: excellent. Translations: the Scaife Viewer carries some translations; Perseus's 2006 CC release covered Greek, Latin AND English collections — but per-text license must be checked (the Hopper FAQ notes publisher agreements restrict some works). Do not bulk-ingest Perseus translations without verifying each text's license flag.

7. Latin — GO (proven end-to-end in pilot)

Open Greek and Latin: GO, same CC BY-SA 4.0 terms as Greek (17M words of Latin). PerseusDL/canonical-latinLit on GitHub contains parallel editions: e.g. data/phi0448/phi001/ (Caesar) ships BOTH phi0448.phi001.perseus-lat2.xml (Holmes Latin edition) AND phi0448.phi001.perseus-eng2.xml (McDevitte & Bohn 1869 English translation, PD by age) in matching CTS book/chapter structure — 404 aligned chapters across 8 books, verified in the pilot. Unicode Latin: excellent (no transliteration needed for alphabetic scripts; schema uses null).

Pilot results (pilot.jsonl)

  • 485 rows, 0 duplicates (SHA-256 of script+translation), ids ancient_pilot_000001…485
  • Chinese (Traditional): 81 — Tao Te Ching, Wang Bi recension + Legge 1891, chapter-aligned
  • Latin: 404 — Caesar De bello Gallico 8 books, Holmes Latin + McDevitte/Bohn English, CTS chapter-aligned
  • Spot checks: TTC ch.1 (道可道…) + pinyin + Legge all correct; Gallic War 1.1 ("Gallia est omnis divisa in partes tres…") correctly paired.

Schema notes / problems hit

  1. transliteration is null for Latin (alphabetic script needs none). Schema permits null; documented per row.
  2. Chinese transliteration is machine pinyin (pypinyin, Mandarin-based, tone numbers). It is NOT scholarly Classical Chinese romanization — flagged in license/notes. A proper literary-Chinese romanization pass is v2 work.
  3. Legge's paragraph numbers ("1.", "2."…) are preserved inside translation (faithful to source; documented).
  4. Sanskrit IAST was NOT included in the pilot (Gita dropped: Arnold unalignable). IAST generation from Devanagari is deterministic — ready when a verse-numbered PD translation is sourced.
  5. Cuneiform rows are deferred to v2: ORACC access was flaky from the sandbox and true Unicode cuneiform needs an ATF→Unicode sign-mapping step. ORACC (CC BY-SA) remains the designated source.
  6. Egyptian rows deferred: TLA unsuitable (German translations, no bulk API). Next step: verify St. Andrews Corpus license.

Recommended build order for the full 25k

  1. Latin (OGL): thousands of chapters available now, proven pipeline.
  2. Greek (OGL): same pipeline; verify translation licenses per text.
  3. Classical Chinese (Wikisource + Gutenberg PD translations): chapter-aligned works (Tao Te Ching, Analects, Mencius, etc.).
  4. Sanskrit (Wikisource Devanagari + verse-numbered PD translation — source TBD).
  5. Cuneiform via ORACC (needs reliable network path + ATF→Unicode mapping).
  6. Egyptian (pending source with English translations + bulk access).