sensix-zo/nllb-paite-600m-v15

NLLB Paite v15: State-of-the-Art English-Paite Translation

This is a highly optimized, state-of-the-art bidirectional translation model for English ↔ Paite (pck). It is fine-tuned from Meta's NLLB (No Language Left Behind) architecture.

This model (v15) represents a significant milestone in low-resource language translation, achieving commercial-grade BLEU scores by prioritizing extreme data quality, contextual nuance, and conversational flow over raw noisy data volume.

Official Benchmarks

Evaluated on a completely unseen, rigorous 5,000-sentence validation set:

Direction	BLEU Score	ChrF++ Score	Edit Distance Similarity
Paite to English	48.16	65.23	63.22%
English to Paite	33.01	58.19	63.67%

Note: The model intentionally achieved a low "Exact Match" score (~6-9%). In machine translation, this proves that the model is dynamically generating natural human language with proper synonyms and phrasing, rather than overfitting or directly memorizing the training data.

Dataset and Training Strategy (The "Clean Corpus")

Unlike previous iterations trained on a massive 1.6 million noisy scraped dataset, v15 was trained on a highly curated "Balanced Corpus" of 340,000 lines:

• 200k "Pure" Paite: Strict, formal, high-quality textbook and literature text to establish precise grammar rules.
• 140k Conversational Subtitles: Colloquial, natural dialogue to teach idioms and natural human conversational flow.

Training Specifications:

• Steps: 36,000
• Batch Size: 64
• Max Length: 512 tokens
• Epochs: ~10 (Optimal convergence prior to overfitting)

Key Capabilities

1. Poetry and Metaphors: Successfully translates nuanced English concepts (e.g., translating Robert Frost's "a road less traveled" into natural Paite phrasing rather than literal timber).
2. Cultural Context: Automatically adapts English honorifics (e.g., "Sir") into proper Paite/Zo cultural titles (e.g., "Pu").
3. Complex Grammar: Handles multi-clause sentences, religious abstractions, and complex timelines flawlessly without losing the context of the initial clause.

---

How to Use (Python)

You can easily deploy this model using the transformers library.

# pip install transformers torch
import torch
from transformers import AutoModelForSeq2SeqLM, AutoTokenizer

# 1. Load the Model & Tokenizer
model_name = "sensix-zo/nllb-paite-v15"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSeq2SeqLM.from_pretrained(model_name)

# 2. Setup your text
text = "Two roads diverged in a yellow wood, and I took the one less traveled by."

# 3. Translate: English to Paite
tokenizer.src_lang = "eng_Latn"
inputs = tokenizer(text, return_tensors="pt")

# Generate Translation
translated_tokens = model.generate(
    **inputs, 
    forced_bos_token_id=tokenizer.convert_tokens_to_ids("pai_Latn"),
    max_length=512,
    num_beams=5,
    no_repeat_ngram_size=2,
    repetition_penalty=1.2,
    do_sample=True,
    temperature=1.0
)

translation = tokenizer.batch_decode(translated_tokens, skip_special_tokens=True)[0]
print(f"Paite Translation: {translation}")

Ablation Study: Clean vs. Noisy Data

This model was tested head-to-head against a previous version trained on 1.6 Million noisy/scraped lines (v13).

While the massive noisy dataset provided high resilience to typographical errors and a vast technical vocabulary, v15 (Clean) drastically outperformed it in grammatical flow, idiom translation, and overall BLEU score. By eliminating scraped website metadata and poor translations, v15 performs less like a direct dictionary mapping and more like a fluent, human interpreter.