aoiandroid / nllb200-brx-bodo

NLLB-200 Bodo & Rajasthani (nllb200-brx-raj)

NLLB-200–based translation model focused on Bodo (brx) and Rajasthani (raj) , two low-resource Indic languages. This model extends NLLB-200 coverage for these languages (e.g. via fine-tuning or vocabulary extension) for use in applications that need offline or on-device translation to/from Bodo and Rajasthani.

Model description

• Architecture: Based on No Language Left Behind (NLLB-200) from Meta AI.
• Languages: Primarily Bodo (brx) and Rajasthani (raj) ; typically used with English (en) as a pivot. Bodo uses the Devanagari script ( brx_Deva ); Rajasthani is also commonly written in Devanagari ( raj_Deva ).
• Intended use: Machine translation between supported languages (e.g. English ↔ Bodo, English ↔ Rajasthani, and related pairs).

Supported languages and scripts

Intended uses and limitations

• Intended use: Translation for documents, UI strings, and general text involving Bodo and Rajasthani, including integration in mobile or offline translation apps.
• Limitations: Low-resource languages; quality may vary by domain and style. Prefer human review for critical or formal content. Performance may be lower than high-resource languages supported by the base NLLB-200 model.
• Out of scope: Legal, medical, or safety-critical translation without human review; use as a sole source of ground truth.

Usage

Hugging Face Transformers

from transformers import AutoModelForSeq2SeqLM, AutoTokenizer

model_id = "aoiandroid/nllb200-brx-raj"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForSeq2SeqLM.from_pretrained(model_id)

# Example: English to Bodo
tokenizer.src_lang = "eng_Latn"
tokenizer.tgt_lang = "brx_Deva"
inputs = tokenizer("Hello, how are you?", return_tensors="pt")
outputs = model.generate(**inputs, forced_bos_token_id=tokenizer.convert_tokens_to_ids("brx_Deva"))
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

With pipeline

from transformers import pipeline

pipe = pipeline("translation", model="aoiandroid/nllb200-brx-raj", src_lang="eng_Latn", tgt_lang="brx_Deva")
result = pipe("Welcome to the app.")
print(result)

Training data

The model is derived from NLLB-200 and may use additional or focused data for Bodo and Rajasthani. The base NLLB-200 model was trained on the FLORES-200 evaluation setup and large-scale multilingual corpora.

Evaluation

Evaluation follows the same methodology as NLLB-200 (e.g. FLORES-200 dev/test) where applicable. Metrics include BLEU, spBLEU, and chrF++ for the supported language pairs.

License

This model is shared under the CC-BY-NC-4.0 (Creative Commons Attribution-NonCommercial 4.0) license. Commercial use requires separate permission or a different license from the model authors.

Citation

If you use this model, please cite the original NLLB-200 work:

@article{nllb2022,
  title={No Language Left Behind: Scaling Human-Centered Machine Translation},
  author={{NLLB Team} and others},
  journal={arXiv preprint arXiv:2207.04672},
  year={2022}
}

Acknowledgments

• Meta AI for the NLLB-200 model and FLORES-200.
• Communities and contributors supporting Bodo and Rajasthani language resources.