TuvanGemma

TuvanGemma is a local Russian ↔ Tuvan translation model distributed as a GGUF file for llama.cpp. The Windows portable application is TuvanTranslator-windows-x64-v0.1.0.zip.

The model is a Gemma-derived model. Read and comply with the Gemma Terms of Use included with the package and linked from the desktop repository before using or redistributing it.

Evaluation results

The E3 release was trained for one epoch on 370,902 directional examples and evaluated on 3,824 held-out pairs in each direction:

direction baseline chrF++ TuvanGemma chrF++ baseline BLEU TuvanGemma BLEU
RU → TYV 13.0657 45.4822 1.1298 16.2996
TYV → RU 19.6637 44.7172 2.1083 20.8963

This is a specialized Russian ↔ Tuvan model. Review names, numbers, idioms, and important translations before relying on its output.

Dataset acknowledgment

This model was fine-tuned using Agisight's tyv-rus-200k dataset: Agisight/tyv-rus-200k on Hugging Face. The dataset is attributed under CC BY 4.0. The complete attribution notice is included in the Windows package.

Desktop application

Source code, build instructions, third-party notices, and the complete Gemma terms are available at kungaa/TuvanTranslator on GitHub.

The ZIP is a self-contained Windows x64 package with CUDA, Vulkan, and CPU llama.cpp runtimes, the model, and a fixed WebView2 runtime. Auto mode tries CUDA, then Vulkan, then CPU. Vulkan requires a working Vulkan-capable Windows graphics driver; vulkan-1.dll is intentionally not bundled.

Downloads last month
19
GGUF
Model size
4B params
Architecture
gemma3
Hardware compatibility
Log In to add your hardware

5-bit

Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for kungaa/TuvanGemma

Quantized
(43)
this model