FormosanBank/formosan-mt

FormosanBank Machine Translation

Public parallel corpora for 15 Indigenous Formosan languages aligned with English and Mandarin Chinese. This release was built from the public FormosanBank snapshot, excludes Formosan-Taiwan-Bible-Society-Bibles, uses canonical kindOf="standard" Formosan text, and includes train-only DeepL pivot translations.

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	368,985	321,355	12,656	34,974	294,349
formosan-zh	425,247	376,231	12,241	36,775	7,723
Total	794,232	697,586	24,897	71,749	302,072

All synthetic rows are in train. Validation and test contain only human translations, records classified as sentences by the source structure, and MT-standardized Formosan text.

The English corpus contains substantially more synthetic data because Chinese-only public records were translated into English. Treat synthetic targets as useful augmentation rather than human reference translations.

Languages

Code	Language	English pairs	Chinese pairs
ami	Amis	80,090	85,291
bnn	Bunun	32,009	39,094
ckv	Kavalan	17,518	20,034
dru	Rukai	40,985	48,263
pwn	Paiwan	23,682	31,014
pyu	Puyuma	24,810	27,891
ssf	Thao	12,341	10,476
sxr	Saaroa	10,591	8,154
szy	Sakizaya	13,262	14,082
tao	Tao / Yami	13,692	16,340
tay	Atayal	28,558	38,590
trv	Seediq / Truku	31,135	37,983
tsu	Tsou	10,509	10,960
xnb	Kanakanavu	16,481	21,246
xsy	Saisiyat	13,322	15,829

Schema

The two main files retain the previous nine-column format.

Field	Type	Description
id	integer	Row identifier within the config.
source_lang	string	Formosan language code.
target_lang	string	en or zh.
source_sentence	string	Canonical Formosan MT text.
target_sentence	string	English or Chinese translation.
lang_code	string	Formosan language code retained for compatibility.
dialect	string	Dialect label or unknown.
source	string	Public source path or identifier.
split	string	train, validate, or test.

The files under provenance/ map each id to stable row IDs, source repository commits, XML identifiers, translation origin, domain bucket, row type, and document ID. They also include the complete release statistics and sanitized split-validation reports.

Loading

Each config remains a single CSV for compatibility, so Hugging Face exposes it as a train split. Use the row-level split field to recover the research splits.

from datasets import DatasetDict, load_dataset

all_rows = load_dataset("FormosanBank/formosan-mt", "formosan-en", split="train")

dataset = DatasetDict({
    name: all_rows.filter(lambda row: row["split"] == name)
    for name in ("train", "validate", "test")
})

For the Chinese corpus, replace formosan-en with formosan-zh. Reverse-direction training can swap source_sentence and target_sentence.

Hard Splits

The split builder targets roughly 90% train, 2.5% validation, and 7.5% test per language, with larger evaluation sets where similarity groups require them. Every language exceeds the configured evaluation floors.

The independent validators report zero train-versus-evaluation conflicts for:

• Exact normalized Formosan, target, and pair text.
• Punctuation-skeleton Formosan, target, and pair text.
• One-character-edit Formosan and target neighbors.
• Character n-gram neighbors at Jaccard similarity 0.82 or above.
• Source-document overlap.

Rows structurally classified as lexemes or non-sentences, ambiguous MT normalization, and synthetic translations are excluded from validation and test. Some sentence-classified XML units remain short or phrase-like. Validation and test are also separated by the same checks within each Formosan language. Identical target phrases can still occur across different Formosan languages.

TAME-MT Exposure Audit

TAME-MT 0.2.2 compared every validation and test sentence with the training set using exact character 3-5 gram retrieval conditioned by Formosan language. The release gate requires zero exact overlap and zero source, target, or pair exposure at 0.95 or above.

Direction	Eval rows	TM BLEU	TM chrF2	Mean source exposure	Source >= 0.70	Source >= 0.95
Formosan to English	47,630	3.80	17.09	0.269	1.153%	0%
English to Formosan	47,630	2.66	17.41	0.253	0.928%	0%
Formosan to Chinese	49,016	0.31	4.41	0.217	0.273%	0%
Chinese to Formosan	49,016	1.47	13.84	0.076	0.118%	0%

TM BLEU and TM chrF2 score nearest-neighbor translation-memory retrieval, not a trained model. Low values indicate that the held-out references cannot generally be recovered by copying the closest training pair. Full reports are under provenance/exposure_*.json.

Provenance

This release was generated by formosan-mt-corpus-v3 from:

• FormosanBank public source commit: c59fb65941217c531a22dc0493af878c68374773
• FormosanBank QC commit: c6c1d06010f54e80e08b14ab2aa8099c3e2284eb
• MT standardization profile: formosan-mt-standard-v3
• Toolkit build commit: 9fb064c7a836ef01351e5f490662f66966bee0b7
• Source build completed: 2026-08-09

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

Intended Use

• Low-resource Formosan machine translation research.
• Formosan to English or Chinese training and evaluation.
• English or Chinese to Formosan training by reversing pairs.
• Dialect-aware and source-domain experiments.

Limitations

• Sources vary in dialect, genre, translation style, and transcription quality.
• Some sentence-classified held-out records are short phrases where longer human data is scarce.
• The English training split is dominated by synthetic DeepL targets.
• Hard similarity filters reduce known leakage but cannot prove semantic independence.
• Automatic metrics are not substitutes for fluent-speaker evaluation.
• Community-facing or high-stakes translations require human review.

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}
}

Terms

Noncommercial research, education, cultural, and language revitalization uses are permitted subject to the FormosanBank terms and source-specific restrictions.

• Terms of Use
• AI Use Addendum