README.md

---
license: cc-by-4.0
language:
- en
- az
base_model: facebook/nllb-200-3.3B
pipeline_tag: translation
tags:
- translation
- azerbaijani
- nllb
- english-azerbaijani
library_name: transformers
datasets:
- LocalDoc/english-azerbaijani-parallel-corpus
---

# NLLB-3.3B English→Azerbaijani

A fine-tuned [NLLB-200-3.3B](https://huggingface.co/facebook/nllb-200-3.3B) model for **English→Azerbaijani** translation. It was trained with LoRA on a ~1.4M-pair corpus combining high-quality distilled translations across multiple domains (formal/encyclopedic, instruction-style, dialogue, and conversational registers), then merged into a self-contained model.

On a 1,012-sentence FLORES-based benchmark it **outperforms Google Translate and several commercial LLM APIs** on English→Azerbaijani, and approaches the strongest proprietary models — at a fraction of their size and cost.

## Benchmark Results

Evaluated on [`LocalDoc/en_az_translate_benchmark`](https://huggingface.co/datasets/LocalDoc/en_az_translate_benchmark) (1,012 sentences, EN→AZ). All metrics are reference-based: chrF++ (`word_order=2`), BLEU (sacreBLEU), and COMET-DA (`Unbabel/wmt22-comet-da`). Higher is better.

| Model | chrF++ | BLEU | COMET-DA |
|---|---|---|---|
| GPT-5.4-mini | 70.08 | 45.61 | 92.86 |
| Gemini-2.5-flash | 69.61 | 45.71 | 92.70 |
| **This model (NLLB-3.3B EN→AZ)** | **69.30** | **44.42** | **92.70** |
| DeepSeek-V4-Pro | 68.67 | 43.88 | 92.78 |
| DeepSeek-V4-Flash | 67.96 | 42.82 | 92.58 |
| GPT-4.1 | 67.76 | 43.03 | 92.71 |
| Google Translate | 66.90 | 41.64 | 92.37 |
| Gemma-4-31B-it | 66.22 | 40.46 | 92.40 |
| GPT-5.4-nano | 62.10 | 33.87 | 91.41 |
| Qwen3.6-35B-A3B | 60.39 | 33.57 | 91.23 |
| NLLB-200-3.3B (base, zero-shot) | 59.03 | 31.76 | 89.86 |

Fine-tuning improved the base NLLB-3.3B by **+10.3 chrF++** (59.03 → 69.30) and **+2.84 COMET-DA** (89.86 → 92.70). The result surpasses Google Translate, DeepSeek-V4-Pro, GPT-4.1, Gemma-3-31B, and Qwen-35B, and comes within 0.8 chrF++ of GPT-5.4-mini and Gemini-2.5-flash.

## Example Translations

A few cases where this model produces more accurate or more natural Azerbaijani than the base NLLB-3.3B and/or Google Translate. Full sentences shown.


**EN:** *No worries, take your time. There's really no rush at all.* — idiomatic, not literal
- **This model:** Narahat olmayın, tələsməyin. Həqiqətən tələsmək lazım deyil.
- **Google:** Narahat olmayın, tələsməyin. Əslində heç bir tələskənlik yoxdur. — literal and stilted

**EN:** *Explain like I'm five: how does the internet actually work?* — natural idiom vs. base
- **This model:** Beş yaşım varmış kimi izah edin: internet əslində necə işləyir? 
- **Base NLLB-3.3B:** Beş yaşında kimi izah edin: İnternet həqiqətən necə işləyir? — *Beş yaşında kimi* is ungrammatical

## Usage

```python
import torch
from transformers import AutoModelForSeq2SeqLM, NllbTokenizer

MODEL = "LocalDoc/nllb-3.3b-en-az"
SRC, TGT = "eng_Latn", "azj_Latn"

tokenizer = NllbTokenizer.from_pretrained(MODEL, src_lang=SRC, tgt_lang=TGT)
model = AutoModelForSeq2SeqLM.from_pretrained(MODEL, torch_dtype=torch.bfloat16).eval()
bos = tokenizer.convert_tokens_to_ids(TGT)

def translate(texts, num_beams=4, max_length=256):
    if isinstance(texts, str):
        texts = [texts]
    tokenizer.src_lang = SRC
    enc = tokenizer(texts, return_tensors="pt", padding=True,
                    truncation=True, max_length=max_length)
    with torch.no_grad():
        gen = model.generate(**enc, forced_bos_token_id=bos,
                             num_beams=num_beams, max_length=max_length)
    return tokenizer.batch_decode(gen, skip_special_tokens=True)

print(translate("The agreement is expected to be signed by the end of the month."))
```

For best results, translate one sentence at a time (the model is sentence-level). Split long texts into sentences before translating.

## Training Details

- **Base model:** facebook/nllb-200-3.3B
- **Method:** LoRA (r=32, alpha=64) on attention and FFN projections, then merged
- **Training data:** ~1.4M EN→AZ pairs, distilled and filtered across domains:
  - Formal / encyclopedic / news
  - Instruction-style (assistant tasks, Q&A)
  - Dialogue and conversational speech
- **Direction:** English → Azerbaijani (direct, no pivot language)
- **Sequence length:** 256 tokens

## Limitations

- **Sentence-level:** translate sentence by sentence; long documents should be split first.
- **Direction:** trained for English→Azerbaijani only.
- **Rare lexical gaps:** very specialized vocabulary (e.g. exotic culinary terms) may occasionally be less precise than large general-purpose systems.
- **Latin script:** outputs standard literary Azerbaijani in Latin script (`azj_Latn`).

## Citation

If you use this model, please cite the LocalDoc organization on Hugging Face.