mbart-mizo-merged

🇮🇳 NLLB-200-Distilled-600M (English ↔ Mizo) — QLoRA Fine-tune

Author: flt007  Base: facebook/nllb-200-distilled-600M

Low-resource English ↔ Mizo translation model fine-tuned with QLoRA.

---

🧠 Overview

A lightweight NLLB derivative trained on bilingual Mizo–English text to improve coverage for the under-represented Mizo language (lus_Latn).

---

⚙️ Usage

from transformers import AutoTokenizer, AutoModelForSeq2SeqLM

model = AutoModelForSeq2SeqLM.from_pretrained('flt007/mbart-mizo-merged')
tok = AutoTokenizer.from_pretrained('flt007/mbart-mizo-merged')

text = 'We must protect our forests.'
inp = tok(text, return_tensors='pt')
out = model.generate(
    **inp,
    forced_bos_token_id=tok.convert_tokens_to_ids('lus_Latn'),
    max_new_tokens=50
)
print(tok.decode(out[0], skip_special_tokens=True))

---

🏗 Training Details

• Method: QLoRA (8-bit LoRA adapters)
• Dataset: Prototype (4 bilingual pairs)
• Epochs: 1
• Hardware: Google Colab T4 GPU (16 GB VRAM)

---

🔮 Next Steps

• Expand dataset to >10 000 pairs
• Multi-epoch training & BLEU/ChrF evaluation
• Bidirectional English ↔ Mizo model release

---

📜 License

Released under CC-BY-NC 4.0

---

❤️ Acknowledgments

• Meta AI (NLLB)
• Hugging Face
• Frankie Thiak