TuvanGemma

TuvanGemma is a family of local Russian ↔ Tuvan translation models. The recommended release fine-tunes google/translategemma-12b-it and is provided as a llama.cpp GGUF, a LoRA adapter, and a self-contained Windows application.

The models are Gemma-derived. Read and comply with the applicable Gemma Terms of Use before using or redistributing them.

Downloads

Release	Download	Notes
12B model (recommended)	tuvangemma-12b-q4_k_m.gguf	Q4_K_M, 7.31 GB; llama.cpp
12B Windows app	TuvanTranslator-windows-x64-v0.2.1-12b.zip	Portable; bundles the model and CUDA, Vulkan, and CPU runtimes
12B LoRA adapter	translategemma-12b-qlora-r16	Use with google/translategemma-12b-it
4B model (smaller legacy release)	tuvangemma-4b-q5_k_m.gguf	Q5_K_M, 2.84 GB
4B Windows app	TuvanTranslator-windows-x64-v0.1.0.zip	Older, lower-memory portable app
4B LoRA adapter	gemma4-e4b-qlora	Adapter used for the legacy release

The v0.2.1 desktop app adds a Lower memory (2K context) switch. Leave it off for the normal 4,096-token context; turn it on before loading the model on an 8 GB GPU. Changing it restarts the local inference runtime automatically.

12B evaluation

The 12B adapter was evaluated on 256 held-out examples in each direction (512 translations total):

Direction	Examples	BLEU	chrF++
Russian → Tuvan	256	20.4171	49.1626
Tuvan → Russian	256	25.9565	49.3654

The evaluation report describes the procedure. Raw predictions are also included under reports/.

These are automatic metrics on one held-out split, not a guarantee of translation quality. Review names, numbers, idioms, and important text with a qualified Tuvan speaker.

GPU memory

Direct llama.cpp Vulkan measurements with full layer offload used 7,853 MiB at 4,096 tokens and 8,113 MiB at 8,192 tokens. In practice:

• 12 GB VRAM: comfortable for the 12B Q4 model at the default context.
• 8 GB VRAM: possible but tight; use the app's 2K context switch and close other GPU-heavy programs. Driver/display overhead varies by machine.
• Less than 8 GB: use partial offload/CPU or the smaller 4B release.

See the VRAM report for the measured configuration and caveats.

Previous 4B evaluation

The earlier 4B release was evaluated separately on 3,824 held-out examples in each direction. Its split and setup differ from the 12B evaluation, so the tables should not be compared as a controlled model-size experiment.

Direction	Baseline chrF++	TuvanGemma chrF++	Baseline BLEU	TuvanGemma BLEU
Russian → Tuvan	13.0657	45.4822	1.1298	16.2996
Tuvan → Russian	19.6637	44.7172	2.1083	20.8963

Dataset acknowledgment

Fine-tuning used Agisight's tyv-rus-200k dataset, attributed under CC BY 4.0. The desktop package includes the dataset attribution notice.

Desktop application

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

The Windows ZIP is self-contained and uses llama.cpp. Auto mode tries CUDA, then Vulkan, then CPU. Vulkan requires a working Vulkan-capable graphics driver; vulkan-1.dll is intentionally supplied by the driver rather than bundled. A Metal backend and macOS packaging recipe are present in the source, but a signed/notarized macOS binary is not currently published.