natgillin/translations

Parquet bitext corpus assembled from upstream mtdata / OPUS releases.

Schema

column	type	description
source	string	source-language sentence
target	string	target-language sentence
source_lang	string	source language, ISO 639-3 (3-char), e.g. bel, cat, eng
target_lang	string	target language, ISO 639-3 (3-char)
origin	string	upstream sub-corpus, e.g. opus-ccmatrix, opus-nllb, opus-gnome, elr, mtdata-merged

Source / target ordering is fixed by alphabetical ISO 639-3 code (bel before eng, cat before eng). This keeps the pair name deterministic and bidirectional joins unambiguous.

Configs / layout

One config per language pair. The OPUS sub-corpus is encoded as a row-level column (origin), not as a separate config.

data/
  cat-eng/
    opus-ccmatrix-00000-of-00200.parquet
    opus-nllb-00000-of-00200.parquet
    opus-gnome-<file_hash>-00000.parquet
    mtdata-merged-00000-of-00200.parquet
    ...
  bel-eng/
    ...

Every pair is pinned to split: train so that filenames containing substrings like -test- (which occur in some ELRC sub-corpora) aren't auto-detected as a test split.

Origins

The origin column carries the upstream sub-corpus identifier. Common values:

• mtdata-merged — mtdata's deduplicated merged bitext for the pair
• opus-<source> — one upstream OPUS sub-corpus, e.g. opus-ccmatrix, opus-nllb, opus-ccaligned, opus-multiccaligned, opus-wikimatrix, opus-eubookshop, opus-kde4, opus-gnome, opus-neulab-tedtalks, opus-elrc-*, opus-tatoeba, opus-ted2020, …
• elr — ELRA / ELR slice of the pair when distinct from the merged top-level
• unknown — rows whose upstream URL didn't match a known OPUS pattern

License

Inherits per-corpus terms from the upstream OPUS sources. Cite OPUS: https://opus.nlpl.eu.