FormosanBank/formosan-mt

FormosanBank Machine Translation

Public parallel corpora for 15 Indigenous Formosan languages aligned to English and Mandarin Chinese. The dataset is published in the same CSV format as earlier releases, but the train/validate/test labels have been rebuilt as leakage-controlled hard splits for more realistic MT evaluation.

The two files are:

• formosan_en_hf.csv: Formosan -> English rows where an English translation is available.
• formosan_zh_hf.csv: Formosan -> Chinese rows where a Chinese translation is available.

Some public rows have Chinese but no English translation, so the Chinese config is larger than the English config.

Important Note

FormosanBank MT models may use this public data together with additional private or restricted corpora. This dataset is the public MT corpus only.

Dataset Summary

Config	Rows	Train	Validate	Test
formosan-en	87,427	83,249	1,480	2,698
formosan-zh	270,082	241,243	7,115	21,724
Total	357,509	324,492	8,595	24,422

The source file used for this release was the public combined corpus with columns:

lang_code,formosan_sentence,chinese_sentence,english_sentence,source,dialect,split

The original split column was ignored and rebuilt from scratch.

What Changed In This Release

The current release uses an in_domain_hard split strategy modeled after the FormosanBank MT experiment splits:

• It removes rows that would create exact normalized leakage between train and validate/test.
• It enforces zero normalized train-vs-eval overlap for Formosan text, target text, and Formosan-target pairs.
• It holds out source documents for validate/test, so evaluation is not just memorized neighboring rows from the same source file.
• It keeps lexical, classroom, dictionary, and other short/easy rows in training where they can help vocabulary learning.
• It excludes short/easy rows from validate/test to avoid inflated scores from dictionary-style examples.
• Validate/test rows require at least 4 Formosan tokens and 4 target-side tokens.

This means the held-out sets are intentionally harder than older random or lightly shuffled splits. Scores on this dataset should be treated as a more honest estimate of out-of-sample MT performance.

Languages

Code	Language	Formosan->English	Formosan->Chinese	Total
ami	Amis	10,183	25,481	35,664
bnn	Bunun	10,353	26,820	37,173
ckv	Kavalan	3,893	15,079	18,972
dru	Rukai	12,833	33,486	46,319
pwn	Paiwan	5,833	20,018	25,851
pyu	Puyuma	6,313	22,064	28,377
ssf	Thao	1,803	9,116	10,919
sxr	Saaroa	1,799	7,301	9,100
szy	Sakizaya	2,668	10,286	12,954
tao	Tao / Yami	2,418	11,133	13,551
tay	Atayal	11,362	29,460	40,822
trv	Seediq / Truku	8,500	28,583	37,083
tsu	Tsou	2,583	7,617	10,200
xnb	Kanakanavu	4,241	13,976	18,217
xsy	Saisiyat	2,645	9,662	12,307

Schema

Both CSV files use the same 9-column schema:

Field	Type	Description
id	integer	Row identifier within the file.
source_lang	string	Formosan source language code, e.g. ami, bnn, tay.
target_lang	string	en for English or zh for Chinese.
source_sentence	string	Formosan sentence or phrase.
target_sentence	string	English or Chinese translation.
lang_code	string	Same Formosan code as source_lang; retained for compatibility.
dialect	string	Dialect label when available, otherwise UNKNOWN.
source	string	Provenance path or source identifier from the upstream corpus.
split	string	One of train, validate, or test.

Loading With datasets

The files are configured as two dataset configs. Because each config is a single CSV file, Hugging Face datasets exposes the file as a train split; use the row-level split column to recover train/validate/test subsets.

from datasets import DatasetDict, load_dataset

repo_id = "FormosanBank/formosan-mt"

ds_en_all = load_dataset(repo_id, "formosan-en", split="train")
ds_zh_all = load_dataset(repo_id, "formosan-zh", split="train")

def split_by_column(ds):
    return DatasetDict({
        "train": ds.filter(lambda ex: ex["split"] == "train"),
        "validate": ds.filter(lambda ex: ex["split"] == "validate"),
        "test": ds.filter(lambda ex: ex["split"] == "test"),
    })

en_splits = split_by_column(ds_en_all)
zh_splits = split_by_column(ds_zh_all)

Filtering To One Language

ami_en = ds_en_all.filter(lambda ex: ex["lang_code"] == "ami")
ami_en_splits = split_by_column(ami_en)

atayal_zh = ds_zh_all.filter(lambda ex: ex["lang_code"] == "tay")
atayal_zh_splits = split_by_column(atayal_zh)

Training Format Example

Most seq2seq MT trainers expect source and target text columns. For Formosan -> English:

example = ds_en_all[0]
source_text = example["source_sentence"]
target_text = example["target_sentence"]
source_lang = example["source_lang"]
target_lang = example["target_lang"]

To train a reverse direction such as English -> Formosan, swap source_sentence and target_sentence in your preprocessing and use lang_code to choose the Formosan target language.

Split Validation

The generated hard splits were validated with these checks:

Config	Train-vs-eval Formosan overlap	Train-vs-eval target overlap	Train-vs-eval pair overlap	Eval short/easy rows
formosan-en	0	0	0	0
formosan-zh	0	0	0	0

The validation and test splits are sentence-like evaluation sets, while lexical/easy examples remain available for training.

Intended Use

• Research on low-resource Formosan machine translation.
• Training and evaluating Formosan -> English and Formosan -> Chinese MT systems.
• Building reverse-direction systems by swapping source and target during preprocessing.
• Dialect-aware and source-domain-aware MT experiments using dialect and source metadata.

Limitations

• The corpus contains heterogeneous sources, dialects, sentence lengths, and translation styles.
• Some rows are short phrase or lexicon-like entries; these are useful for training but excluded from validate/test in this hard split.
• The split prevents exact normalized leakage, but it cannot guarantee semantic non-overlap between paraphrases.
• English and Chinese coverage differ because not every public row has both translations.
• Community-facing or high-stakes translation systems should include fluent-speaker review.

Citation

If you use this dataset, please cite FormosanBank and the upstream corpus sources where applicable.

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