PyaeSoneK/nllb-600m-wikihow-en-my

NLLB-200-distilled-600M fine-tuned for English to Burmese (WikiHow-MY)

This is facebook/nllb-200-distilled-600M fine-tuned for English to Burmese (eng_Latn to mya_Mymr) translation on an instructional-text corpus derived from wikiHow. It targets the procedural / how-to register, where the zero-shot NLLB baseline is weakest.

Model description

• Base model: facebook/nllb-200-distilled-600M (600M-param distilled NLLB-200)
• Architecture: M2M100ForConditionalGeneration (seq2seq, SentencePiece tokenizer)
• Direction: English (eng_Latn) to Burmese (mya_Mymr), single direction
• Fine-tuning data: WikiHow-MY instructional EN to MY pairs (see Training data)
• Selection metric: dev chrF (more stable than BLEU on unsegmented Burmese)

Intended uses & limitations

Intended use. Translating English instructional / how-to text into Burmese (steps, tips, procedural prose). Research and non-commercial use only.

Out of scope / limitations.

• Non-commercial only — inherits CC-BY-NC-4.0 from NLLB-200 (see License).
• Tuned on a single domain (wikiHow). Expect degradation on conversational, legal, medical, or other out-of-domain text.
• Single direction (en to my). Do not use for my to en.
• Burmese has no orthographic word boundaries; downstream metrics and any word-level processing must account for this. Reported BLEU is spBLEU (SentencePiece-tokenized) and the primary metric is chrF++.
• May hallucinate, drop, or mistranslate named entities and numbers; not for high-stakes use without human review.

Training data

Fine-tuned on WikiHow-MY, parallel English to Burmese instructional sentence pairs extracted from wikiHow articles. wikiHow content is licensed CC-BY-NC-SA-3.0; this derivative therefore carries a non-commercial, share-alike obligation in addition to NLLB's CC-BY-NC (see License & attribution).

• Train / dev / test: 8,302 / 908 / 846 pairs (~10,056 total, from 82 articles)
• Article-disjoint splits (seed 42): no article appears in more than one split, so the benchmark has no train/test leakage at the article level.
• Preprocessing: Zawgyi-to-Unicode normalization, de-duplication, and sentence-level alignment of the English-Burmese instructional pairs.

Training procedure

Fine-tuned with 🤗 Transformers Seq2SeqTrainer.

Hyperparameter	Value
base model	facebook/nllb-200-distilled-600M
learning rate	3e-5
warmup ratio	0.05
epochs	up to 10 (early stop on dev chrF)
per-device train batch	4
grad accumulation	8 (effective batch 32)
precision	fp16
eval / save steps	250
early-stopping patience	4
max length	256
seed	42
beams (eval/inference)	5

Evaluation results

Evaluated on FLORES-200 devtest (eng_Latn-mya_Mymr) and the held-out WikiHow-MY test split. chrF++ and spBLEU computed with sacrebleu.

System / test set	chrF++	spBLEU	BLEU
This model — WikiHow-MY test	41.64	23.18	3.74
Base NLLB zero-shot — WikiHow-MY	36.01	19.33	2.62
This model — FLORES-200 devtest	33.50	17.78	2.71
Base NLLB zero-shot — FLORES-200	29.17	14.79	2.19

Fine-tuning on WikiHow improves the procedural in-domain test by +5.6 chrF++ (36.01 to 41.64) and also lifts general-domain FLORES-200 by +4.3 chrF++ (29.17 to 33.50) — i.e. domain adaptation with no catastrophic forgetting. chrF++ is the primary metric (segmentation-agnostic for unsegmented Burmese); spBLEU uses the FLORES-200 tokenizer. Numbers from experiments/results/main_results.json.

How to use

from transformers import AutoModelForSeq2SeqLM, AutoTokenizer

model_id = "PyaeSoneK/nllb-600m-wikihow-en-my"
tokenizer = AutoTokenizer.from_pretrained(model_id, src_lang="eng_Latn")
model = AutoModelForSeq2SeqLM.from_pretrained(model_id)

text = "Fold the paper in half, then crease the edge firmly."
inputs = tokenizer(text, return_tensors="pt")

# Force the decoder to start generating in Burmese.
bos = tokenizer.convert_tokens_to_ids("mya_Mymr")
generated = model.generate(
    **inputs,
    forced_bos_token_id=bos,
    num_beams=5,
    max_new_tokens=256,
)
print(tokenizer.batch_decode(generated, skip_special_tokens=True)[0])

Note: on recent transformers, prefer tokenizer.convert_tokens_to_ids("mya_Mymr") over the deprecated tokenizer.lang_code_to_id[...].

License & attribution

This model is released under CC-BY-NC-4.0. Two upstream non-commercial terms apply and you must comply with both:

1. NLLB-200 (facebook/nllb-200-distilled-600M) is CC-BY-NC-4.0. Any derivative — including this fine-tune — must remain non-commercial and attribute Meta AI / the NLLB Team.
2. wikiHow training content is CC-BY-NC-SA-3.0: non-commercial and share-alike. Reuse of this model or its outputs must credit wikiHow and carry a compatible non-commercial license.

Effective terms = the union of these: non-commercial use only, attribution to both NLLB and wikiHow, and share-alike where wikiHow-derived content is redistributed.

Citation

@misc{nllb600m_wikihow_en_my,
  title  = {NLLB-200-distilled-600M fine-tuned for English-Burmese on WikiHow-MY},
  author = {Pyae Sone Kyaw},
  year   = {2026},
  note   = {Fine-tune of facebook/nllb-200-distilled-600M, CC-BY-NC-4.0}
}

@article{nllb2022,
  title   = {No Language Left Behind: Scaling Human-Centered Machine Translation},
  author  = {{NLLB Team} and Costa-juss\`a, Marta R. and others},
  journal = {arXiv preprint arXiv:2207.04672},
  year    = {2022}
}