A fine-tuned 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 (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
This model (NLLB-3.3B EN→AZ)	70.00	45.86	92.75
Gemini-2.5-flash	69.61	45.71	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 +10.97 chrF++ (59.03 → 70.00) and +2.89 COMET-DA (89.86 → 92.75). 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.