---
title: NogaiLLM Russian-Nogai (experimental)
emoji: 😻
colorFrom: yellow
colorTo: pink
sdk: gradio
sdk_version: 6.19.0
python_version: '3.13'
app_file: app.py
pinned: false
license: apache-2.0
short_description: Experimental Russian-Nogai translation (Qwen2.5-1.5B LoRA)
models:
- Qwen/Qwen2.5-1.5B-Instruct
- ansarzeinulla/Qwen2.5-1.5B-Nogai-LoRA
- ansarzeinulla/Qwen2.5-1.5B-Nogai-SFT-Experimental
datasets:
- ansarzeinulla/Nogai-Unified-Corpus-v1
- ansarzeinulla/Nogai-Russian-SFT-Biblical-v1
---

# NogaiLLM demo: Russian ↔ Nogai (experimental)

Chat demo of [NogaiLLM](https://github.com/ansarzeinulla/NogaiLLM-Apple-Silicon): Qwen2.5-1.5B-Instruct adapted to Nogai in two LoRA stages trained with `mlx-lm` on an Apple M2 Pro.

## How it works

1. Downloads both MLX adapters and converts them to PEFT format: all seven target modules (q/k/v/o/gate/up/down), layers 12–27.
2. Merges the **Phase 1** adapter ([Qwen2.5-1.5B-Nogai-LoRA](https://huggingface.co/ansarzeinulla/Qwen2.5-1.5B-Nogai-LoRA), continued pre-training on Nogai text) into Qwen2.5-1.5B-Instruct.
3. Applies the **Phase 2** adapter ([Qwen2.5-1.5B-Nogai-SFT-Experimental](https://huggingface.co/ansarzeinulla/Qwen2.5-1.5B-Nogai-SFT-Experimental), translation instructions) on top.
4. Generates on CPU in float32, temperature 0.3, up to 128 new tokens.

## Try

Use the format the model was trained on:

- `Переведи этот текст на ногайский язык: В начале сотворил Бог небо и землю.`
- `Переведи этот текст на русский язык: Бизикилер, яшавлары пайдасыз болмасын.`

## Limitations

- Trained on Bible text only. Modern-domain sentences often come back with religious vocabulary, or with words from other languages.
- Free CPU hardware: the first reply after the Space wakes up takes a while.
- No quality numbers yet. A clean evaluation (chrF++/BLEU against human translations on a held-out test set) is in progress.