Yasalma

Community-built infrastructure for Tatar AI

News

  • Whisper large v3 Tatar–02/10/2026

    openai/whisper-large-v3, fully fine-tuned on Common Voice Scripted Speech 27.0 Tatar (CC0). It writes Tatar in Cyrillic, with capital letters and punctuation, so its output can be used as subtitles with little editing.

  • imlaTAT 0.7–29/09/2026

    OsmanLat (Ottoman to Latin Turkish model) retrained on 3+k lines from seven printed editions of 1906 to 1935. Reads Arabic-script Old Tatar and outputs Cyrillic. On held-out lines the character error is 3.4 % and the word error is 17 %. With the lexicon decoder they are 2.9 % and 12.6 %. Runs offline, 0.1 to 0.2 s per line.

  • TATlit–31/08/2026

    TATlit is a 478M-parameter, from-scratch base language model for literary Tatar. Pretrained on Tatar and Qypchaq siblings, then solely on Tatar. On FLORES, BOUQuET and TatBLiMP it scores best among the tested models, from Tatar-specialized ones to frontier models of up to 120B parameters.

All news

Models

Data