nllb200-zh-formosan-spm8k

Direction: Traditional Chinese to Formosan
Base model: facebook/nllb-200-distilled-600M
Recipe: nllb200-spm8k-directional-v5
Release: 20260816-193900, validation-selected step 300,000

This is a directional model for 15 Formosan languages. It uses the private_no_bible leakage-controlled corpus, Formosan-aware 8k SentencePiece extension, balanced language/source sampling, direction/language controls, fixed coarse domain tags, and dialect tags. Domain and dialect metadata use independent training dropout so unknown and default are learned inference conditions. The model weights are public. The training corpus is distributed separately to authorized FormosanBank members through the access-controlled FormosanBank/formosan-mt-private dataset and is not included with the weights.

Model details

Item	Value
Base revision	f8d333a098d19b4fd9a8b18f94170487ad3f821d
Training rows	625,460
Effective batch size	64
Maximum sequence length	384
Learning rate	2e-05
Precision	bf16
Checkpoint selection	Validation chrF2
Formosan text	kindOf=standard, formosan-mt-standard-v3
Corpus SHA-256	0024e7fcd35958455a2a060b20999e94b5e8a4364e7f54ba97ef27810746d932
Training profile SHA-256	4253b1b136fedebc589da942b80df71945212d1f4b92a8e98ec6fb160712f14b

Usage

import torch
from transformers import AutoModelForSeq2SeqLM, NllbTokenizer


model_id = "FormosanBank/nllb200-zh-formosan-spm8k"
tokenizer = NllbTokenizer.from_pretrained(model_id, use_fast=False)
model = AutoModelForSeq2SeqLM.from_pretrained(model_id)
model.to("cuda" if torch.cuda.is_available() else "cpu")
NLLB_LIDS = {'ami': 'ami_Latn', 'bnn': 'bnn_Latn', 'ckv': 'ckv_Latn', 'dru': 'dru_Latn', 'pwn': 'pwn_Latn', 'pyu': 'pyu_Latn', 'ssf': 'ssf_Latn', 'sxr': 'sxr_Latn', 'szy': 'szy_Latn', 'tao': 'tao_Latn', 'tay': 'tay_Latn', 'trv': 'trv_Latn', 'tsu': 'tsu_Latn', 'xnb': 'xnb_Latn', 'xsy': 'xsy_Latn'}

def translate(text, lang_code, source_bucket="unknown", dialect="default"):

    tokenizer.src_lang = 'zho_Hant'
    source_bucket = source_bucket if source_bucket in ('dictionary', 'classroom', 'narrative', 'linguistic', 'education', 'media', 'culture', 'religious', 'unknown') else "unknown"
    domain_tag = f"<dom_{source_bucket}>"
    if tokenizer.convert_tokens_to_ids(domain_tag) == tokenizer.unk_token_id:
        domain_tag = "<dom_unknown>"
    dialect_tag = f"<dialect_{dialect}>"
    if tokenizer.convert_tokens_to_ids(dialect_tag) == tokenizer.unk_token_id:
        dialect_tag = "<dialect_default>"
    prompt = (
        f"<to_{lang_code}> <src_zh> {domain_tag} {dialect_tag} {text}"
    )
    inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
    output = model.generate(
        **inputs,
        decoder_start_token_id=tokenizer.eos_token_id,
        forced_bos_token_id=tokenizer.convert_tokens_to_ids(NLLB_LIDS[lang_code]),
        max_new_tokens=256,
        num_beams=4,
    )
    return tokenizer.batch_decode(output, skip_special_tokens=True)[0]

print(translate('他回家了。', "ami"))

The control tags are part of the training contract. Use unknown and default when source bucket or dialect metadata is unavailable.

Evaluation

The best checkpoint was selected on validation chrF2. Test and validation contain only eligible, human-translated sentence pairs. Synthetic pivots and lexical entries are train-only. The headline result uses default metadata controls, so it does not assume access to test-set domain or dialect labels.

Split	Rows
Train	625,460
Test	73,591
Validate	36,797

Scope	BLEU	chrF2	TER
Hard test	10.52	35.06	80.02
Selection validation	9.46	35.82	83.08

Test empty-output rate: 0.0000%.

Language	Samples	BLEU	chrF2	TER
ami	14,348	8.40	30.97	83.12
bnn	6,048	7.78	36.24	84.72
ckv	2,871	22.62	48.09	60.42
dru	6,588	4.03	28.21	97.14
pwn	5,945	6.09	33.16	90.98
pyu	4,334	11.81	37.37	76.99
ssf	2,150	21.53	50.02	56.10
sxr	1,968	7.19	41.72	86.80
szy	2,300	17.42	43.22	67.77
tao	2,360	12.30	36.10	72.84
tay	7,161	4.62	24.52	92.61
trv	10,728	10.91	33.51	72.50
tsu	2,048	12.91	38.60	73.06
xnb	2,590	18.48	50.69	61.18
xsy	2,152	21.81	47.25	63.76

The corpus gate enforces standard-tier Formosan text, 5/10 validation/test proportions from all deduplicated pairs, capacity-aware source balance, sentence-only evaluation, and zero exact, skeleton, one-edit, or configured high character n-gram train/evaluation conflicts. Document overlap is diagnostic. This release passed all leakage gates: exact 0, skeleton 0, one-edit 0, character n-gram 0. Document overlap: 1439.

See eval/metrics.json for sacreBLEU signatures, per-language, source, dialect, and length diagnostics. publication.json records the corpus, profile, run, and checkpoint hashes used for this release.

Intended use

This model supports research, corpus development, and assisted translation for the 15 included Formosan languages. It is designed for the exact prompt and generation contract shown above.

Limitations

Outputs require knowledgeable speaker review. Aggregate metrics hide large differences among languages and domains. This model is not suitable for authoritative, medical, legal, or safety-critical translation.