EN-MN Human Rights Translator
Fine-tune of facebook/nllb-200-distilled-600M for English-to-Mongolian translation, specialized on human rights terminology.
- Source language code:
eng_Latn - Target language code:
khk_Cyrl(NLLB/FLORES-200 code for Halh Mongolian; note this is notmon_Cyrl, which does not exist in NLLB's vocabulary)
Training data
~4,000 professionally translated EN-MN sentence pairs from human rights reports and legal documents, deduplicated and split 80/10/10 into train/val/test.
Test set results
| Metric | Score |
|---|---|
| BLEU (sacrebleu) | 17.56 |
| chrF | 47.39 |
Usage
from transformers import AutoModelForSeq2SeqLM
from transformers.models.nllb.tokenization_nllb import NllbTokenizer
model_id = "Oyunbat26/mongolian-en-mn-translator"
tokenizer = NllbTokenizer.from_pretrained(model_id, src_lang="eng_Latn", tgt_lang="khk_Cyrl")
model = AutoModelForSeq2SeqLM.from_pretrained(model_id)
model.generation_config.forced_bos_token_id = tokenizer.convert_tokens_to_ids("khk_Cyrl")
inputs = tokenizer("Human rights are universal.", return_tensors="pt")
output = model.generate(**inputs, max_length=128, num_beams=4)
print(tokenizer.batch_decode(output, skip_special_tokens=True)[0])