FormosanBank/formosan-mt

FormosanBank Machine Translation

Public parallel corpora for 15 Indigenous Formosan languages aligned with English and Mandarin Chinese. This release uses canonical MT-standardized Formosan text, excludes Formosan-Taiwan-Bible-Society-Bibles, and keeps DeepL pivot translations in training only.

Commercial AI use is prohibited without prior written permission. See the FormosanBank Terms of Use.

Release Summary

Config	Rows	Train	Validate	Test	Synthetic train
formosan-en	328,943	296,930	10,674	21,339	222,274
formosan-zh	415,361	353,961	20,469	40,931	6,123
Total	744,304	650,891	31,143	62,270	228,397

Human pairs in every language use 70% train, 10% validation, and 20% test for Formosan-English, and 85% train, 5% validation, and 10% test for Formosan-Chinese. Evaluation contains only eligible human sentence references. Synthetic pivots and short entries are added to training, so final augmented-corpus percentages are more train-heavy.

Languages

Code	Language	English pairs	Chinese pairs
ami	Amis	67,287	88,403
bnn	Bunun	27,609	36,112
ckv	Kavalan	15,335	17,410
dru	Rukai	34,711	45,769
pwn	Paiwan	24,518	30,265
pyu	Puyuma	23,226	29,940
ssf	Thao	10,893	12,913
sxr	Saaroa	8,080	10,955
szy	Sakizaya	11,635	13,754
tao	Tao / Yami	11,765	14,064
tay	Atayal	30,581	39,664
trv	Seediq / Truku	28,197	34,653
tsu	Tsou	9,077	10,616
xnb	Kanakanavu	13,647	17,050
xsy	Saisiyat	12,382	13,793

Schema

The two main files preserve the established nine-column format: id, source_lang, target_lang, source_sentence, target_sentence, lang_code, dialect, source, and row-level split.

from datasets import DatasetDict, load_dataset

rows = load_dataset("FormosanBank/formosan-mt", "formosan-en", split="train")
dataset = DatasetDict({
    split: rows.filter(lambda row: row["split"] == split)
    for split in ("train", "validate", "test")
})

Use formosan-zh for Chinese. Reverse-direction training can swap the sentence columns. Files under provenance/ map release IDs to source commits and XML records and include independent validation and TAME-MT reports.

Split Quality

The builder groups exact normalized pairs and punctuation skeletons, blocks one-edit conflicts, and excludes pair exposure at character 3-5 gram Jaccard similarity 0.95 or above. These are row-level hard splits, not document-held-out splits.

Direction	Eval rows	TM BLEU	TM chrF2	Mean source exposure	Source >= 0.70
English to Formosan	32,013	3.70	19.54	0.256	0.943%
Formosan to English	32,013	4.98	19.36	0.258	0.975%
Formosan to Chinese	61,400	1.22	6.42	0.227	0.583%
Chinese to Formosan	61,400	2.53	16.06	0.092	0.510%

TM scores measure nearest-neighbor translation-memory retrieval, not model quality. All four directions have zero exact overlap and zero source, target, or pair exposure at 0.95.

Provenance

• Public FormosanBank commit: 3a3c47c220520113f747e6a2d441494000e13c4b
• FormosanBank QC commit: acc3ec9f3137a59b9661e446a750ee1606720394
• Toolkit commit: 573488b215ee1cf3701519a5bcd1010e8f0c6548
• MT standardization: formosan-mt-standard-v3
• Build completed: 2026-08-24

Artifact hashes and per-language counts are recorded in provenance/release_metadata.json and SHA256SUMS.

Limitations

• Sources vary in dialect, genre, translation style, and transcription quality.
• Some human references retain source or linguistic annotations.
• The English training corpus contains substantial synthetic augmentation.
• Similarity controls do not prove semantic or document independence.
• Automatic metrics do not replace evaluation by fluent speakers.

Citation

@misc{formosanbank_mt_public_v3,
  title        = {FormosanBank Machine Translation Public Corpus},
  author       = {FormosanBank contributors},
  year         = {2026},
  howpublished = {https://huggingface.co/datasets/FormosanBank/formosan-mt}
}

See the Terms of Use and AI Use Addendum.