README.md

---
license: cc-by-nc-4.0
library_name: transformers
pipeline_tag: translation
base_model: facebook/nllb-200-distilled-600M
language:
- ami
- bnn
- ckv
- dru
- pwn
- pyu
- ssf
- sxr
- szy
- tao
- tay
- trv
- tsu
- xnb
- xsy
- en
tags:
- translation
- nllb-200
- formosan-languages
- low-resource
metrics:
- bleu
- chrf
- ter
model-index:
- name: nllb200-formosan-en-spm8k
  results:
  - task:
      type: translation
      name: Translation
    dataset:
      name: FormosanBank private no-Bible hard test
      type: private-no-bible-hard-test
      split: test
    metrics:
    - type: bleu
      name: sacreBLEU
      value: 9.324037
    - type: chrf
      name: chrF2
      value: 26.929850
    - type: ter
      name: TER
      value: 96.958093
---

# nllb200-formosan-en-spm8k

**Direction:** Formosan to English<br>
**Base model:** [`facebook/nllb-200-distilled-600M`](https://huggingface.co/facebook/nllb-200-distilled-600M)<br>
**Recipe:** `nllb200-spm8k-directional-v3`<br>
**Release:** `20260809-210523`, validation-selected step 210,000

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, and direction/domain/dialect control tags. The model
weights are public, but the private training corpus is not included.

## Model details

| Item | Value |
|---|---|
| Base revision | `f8d333a098d19b4fd9a8b18f94170487ad3f821d` |
| Training rows | 573,657 |
| Effective batch size | 64 |
| Maximum sequence length | 384 |
| Learning rate | 2e-05 |
| Precision | `bf16` |
| Checkpoint selection | Human validation `chrF2` |
| Formosan text | `kindOf=standard`, `formosan-mt-standard-v3` |
| Corpus SHA-256 | `e3feeaf7c3c51b9cd5c4b0537ffe44a370e7d02723f34b467479b6c4f5f0ea77` |
| Training profile SHA-256 | `34f45832bdedc1b8322b26936dd88667dfecb360d941a7f54f60ae31a1fb4e10` |

## Usage

```python
import torch
from transformers import AutoModelForSeq2SeqLM, NllbTokenizer
from formosan_mt_inference import normalize_formosan


model_id = "FormosanBank/nllb200-formosan-en-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"):
    text = normalize_formosan(text, lang_code)
    tokenizer.src_lang = NLLB_LIDS[lang_code]
    prompt = (
        f"<to_eng> <src_{lang_code}> <dom_{source_bucket}> "
        f"<dialect_{dialect}> {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('eng_Latn'),
        max_new_tokens=256,
        num_beams=4,
    )
    return tokenizer.batch_decode(output, skip_special_tokens=True)[0]

print(translate("Pa'araw cingra.", "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 human validation chrF2. Test references
are human 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 | 573,657 |
| Test | 63,140 |
| Validate | 20,733 |

| Scope | BLEU | chrF2 | TER |
|---|---:|---:|---:|
| Hard test | 9.32 | 26.93 | 96.96 |
| Selection validation | 9.53 | 29.70 | 92.68 |

Test empty-output rate: 0.0000%.


### Confidence intervals

Stratified bootstrap, 200 samples, 95% confidence.

| Metric | Lower | Upper |
|---|---:|---:|
| BLEU | 9.18 | 9.48 |
| chrF2 | 26.78 | 27.08 |
| TER | 96.48 | 97.54 |


| Language | Samples | BLEU | chrF2 | TER |
|---|---:|---:|---:|---:|
| `ami` | 11,017 | 7.62 | 26.29 | 93.14 |
| `bnn` | 6,049 | 14.68 | 33.73 | 77.16 |
| `ckv` | 2,910 | 10.84 | 30.22 | 96.13 |
| `dru` | 5,846 | 1.07 | 13.93 | 135.78 |
| `pwn` | 5,811 | 7.65 | 24.46 | 106.90 |
| `pyu` | 4,315 | 12.50 | 29.72 | 107.12 |
| `ssf` | 1,630 | 14.05 | 31.21 | 87.89 |
| `sxr` | 2,081 | 11.33 | 26.27 | 96.93 |
| `szy` | 1,852 | 13.59 | 32.77 | 90.81 |
| `tao` | 2,257 | 9.04 | 27.97 | 93.45 |
| `tay` | 6,411 | 10.46 | 26.58 | 104.63 |
| `trv` | 6,502 | 12.81 | 31.71 | 96.50 |
| `tsu` | 2,003 | 2.39 | 19.11 | 95.33 |
| `xnb` | 2,682 | 5.62 | 25.89 | 93.22 |
| `xsy` | 1,774 | 16.10 | 35.31 | 83.12 |

The corpus gate enforces standard-tier Formosan text, at least 7.5% test and
2.5% validation per language, human sentence-only evaluation, and zero exact,
skeleton, one-edit, configured high character n-gram, or document
train/evaluation conflicts. This release passed all gates: exact
0, skeleton
0, one-edit
0, character n-gram
0, and document
0.

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.