Voldis/nllb-1.3b-rus-bak

DevLake: NLLB-200 (1.3B) for Russian-Bashkir Translation

Current Model	Architecture	Focus
🔴 Large (This Model)	NLLB-1.3B (QLoRA)	Best Quality (SOTA)
🟡 Medium Model	M2M-100 (418M)	Balanced
🟢 Small Model	MarianMT (77M)	Fastest / CPU

Model Description

This is the High-Performance model submitted by Team DevLake for the LoResMT 2026 Shared Task. It achieved the highest score in our experiments (52.67 CHRF++), significantly outperforming standard baselines.

It is a fine-tuned version of NLLB-200-1.3B-Distilled, trained using QLoRA (4-bit quantization) on a rigorously filtered subset of the Russian-Bashkir parallel corpus.

• Paper/Code: GitHub Repository
• Developed by: DevLake Team
• Language Pair: Russian (rus_Cyrl) $\to$ Bashkir (bak_Cyrl)

Performance

Model	Size	CHRF++	Note
DevLake NLLB	1.3B	52.67	Best Morphology & Syntax
DevLake M2M	418M	48.80	Good baseline
DevLake Marian	77M	43.15	Fast but hallucinates

Usage

Due to the use of QLoRA, this model requires peft and bitsandbytes.

pip install torch transformers peft bitsandbytes accelerate

import torch
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
from peft import PeftModel

# 1. Load Base Model (NLLB)
base_model_id = "facebook/nllb-200-1.3B"
model = AutoModelForSeq2SeqLM.from_pretrained(
    base_model_id,
    load_in_4bit=True,
    device_map="auto"
)

# 2. Load DevLake Adapters
adapter_model_id = "Voldis/nllb-1.3b-rus-bak"
model = PeftModel.from_pretrained(model, adapter_model_id)
tokenizer = AutoTokenizer.from_pretrained(adapter_model_id)

# 3. Inference
text = "Утром я выпил чашку кофе."
inputs = tokenizer(text, return_tensors="pt").to("cuda")

with torch.no_grad():
    generated_tokens = model.generate(
        **inputs,
        forced_bos_token_id=tokenizer.convert_tokens_to_ids("bak_Cyrl"),
        max_length=128
    )

print(tokenizer.batch_decode(generated_tokens, skip_special_tokens=True)[0])

Training Details

• Filtering: We used a BERT-based metric to select only the top 486k sentence pairs (Semantic Similarity $\ge 0.80$).
• Method: QLoRA (Rank=64, Alpha=64).
• Hardware: Trained on a single NVIDIA RTX 3080.

Citation

@inproceedings{tyurin-2026-devlake,
    title = "{D}ev{L}ake at {L}o{R}es{MT} 2026: The Impact of Pre-training and Model Scale on {R}ussian-{B}ashkir Low-Resource Translation",
    author = "Tyurin, Vyacheslav",
    booktitle = "Proceedings for the Ninth Workshop on Technologies for Machine Translation of Low Resource Languages (LoResMT 2026)",
    month = mar,
    year = "2026",
    address = "Rabat, Morocco",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2026.loresmt-1.18",
    doi = "10.18653/v1/2026.loresmt-1.18",
    pages = "209--212",
}