Chuvash AI

community
Activity Feed

AI & ML interests

None defined yet.

Recent Activity

Organization Card

Чӑваш ИИ · Chuvash AI

Open AI for the Chuvash language — translation models, LLM adaptation and open data

🎯 Mission

Chuvash (чӑваш чӗлхи) is a Turkic language of the Volga region with under a million speakers — too small for big AI companies to care, too important for its community to lose. We adapt open multilingual LLMs to Chuvash with continued pre-training and supervised fine-tuning, and release all data and models openly.

📦 Datasets

DatasetSizeDescription
chuvash-ru-parallel1,459,170 pairsLargest public Chuvash–Russian parallel corpus (CC0)
chuvash-en-parallel187,531 pairsLargest public Chuvash–English parallel corpus (CC0)

🤖 Models

LFM3-1.7B-chv — Chuvash-adapted translation model family, released here with GGUF quants. The full fine-tune v2 is out now — our flagship.

📊 Results (chrF++, held-out real parallel data)

Modelru → chvchv → ru
MarianMT baseline23.3—
LFM3-1.7B-chv v1 (pilot, 2.6% of data)33.934.1
LFM3-1.7B-chv v2 (full FT, all data)34.935.1

🤝 Contribute

  • Data — Chuvash texts, parallel pairs, dictionaries: the main bottleneck
  • Feedback — try the models, report weak spots
  • Ideas — tokenizer surgery, LaBSE cleaning, back-translation, English direction

📮 Contact

Founder: @efi-team

Ырӑ кун! · Chuvash AI · Data: CC0 · Models: LFM Open License