ngaihte/nllb-paite-eng

🌍 NLLB Paite → English Translation Model

This is a fine-tuned version of facebook/nllb-200-distilled-600M, trained on a low-resource parallel dataset of Paite → English sentences. It is built to help translate the indigenous Paite language into English using the NLLB (No Language Left Behind) multilingual framework.

---

📌 Model Details

• Model architecture: Transformer (Seq2Seq)
• Base model: facebook/nllb-200-distilled-600M
• Source language: Paite (paite)
• Target language: English (eng_Latn)
• Tokenizer: NllbTokenizerFast
• Training dataset: Custom parallel corpus (approx. 30,000 lines)
• Trained using: Hugging Face 🤗 Transformers with Seq2SeqTrainer
• Training environment: Google Colab Pro (GPU)
• BLEU score (final epoch): 32.51

---

📈 Training Results

Epoch	Training Loss	Validation Loss	BLEU Score
1	1.4355	1.2799	28.99
2	1.1684	1.1814	31.84
3	1.0246	1.1701	32.51

---

🧪 Intended Use

This model is intended for translation tasks involving Paite → English. It is suitable for researchers, linguists, and developers working with underrepresented or endangered languages.

🔁 For English → Paite, a separate model should be fine-tuned using the reversed corpus.

---

🚀 Usage Example

from transformers import AutoTokenizer, AutoModelForSeq2SeqLM

model = AutoModelForSeq2SeqLM.from_pretrained("your-username/nllb-paite-eng")
tokenizer = AutoTokenizer.from_pretrained("your-username/nllb-paite-eng")

source_lang = "paite"
target_lang = "eng_Latn"
text = "Na damzel hia?"

# Add the language tag before encoding
inputs = tokenizer(f">>{source_lang}<< {text}", return_tensors="pt")
translated_tokens = model.generate(
    **inputs,
    forced_bos_token_id=tokenizer.convert_tokens_to_ids(f">>{target_lang}<<")
)

output = tokenizer.decode(translated_tokens[0], skip_special_tokens=True)
print("Translation:", output)