TATlit is a 478M-parameter, from-scratch base language model for literary Tatar (ISO 639-3 tat), a Qypchaq Turkic language written in Cyrillic. It is trained without adapting any English- or Russian-dominant base, so that it produces native Tatar morphosyntax by default rather than acting as a translation layer over a dominant language. This is a base model, not instruction-tuned.

Usage

python

# pip install sentencepiece
from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("yasalma/TATlit")
tok   = AutoTokenizer.from_pretrained("yasalma/TATlit", use_fast=False)

The tokenizer is SentencePiece-based; pass use_fast=False so tokenization matches training exactly (the auto-converted fast tokenizer differs on the leading token). The weights are a standard Llama-architecture checkpoint in safetensors.

Model

Llama-style decoder, 24 layers, hidden size 1280, 16 attention heads with 4-way grouped-query attention, SwiGLU, RoPE, RMSNorm, tied embeddings, context 4096.
48K SentencePiece Unigram tokenizer, Tatar fertility 4.2-4.5 chars/token, 1.5-3x more efficient than general tokenizers.
478M parameters, bf16.

Training

Two stages. Stage 1 pretrains on a Cyrillic-Qypchaq pool (Tatar 43 / Kazakh 30 / Kyrgyz 16 / Bashkir 10 / smaller siblings ~1 percent) under target-capped sampling. Stage 2 specializes on pure, register-weighted literary Tatar with an anti-memorization regime that keeps verbatim reproduction low.

Results

TATlit is compared against three Tatar-specialized baselines (Goldfish-tat 125M, Tweety-7B-tatar, mGPT-1.3B-tatar) and four frontier models from 31B to 120B (Gemma-4-31B, Llama-4-Scout, Qwen3-32B, gpt-oss-120b), each measured with the same harness per benchmark.

Byte-per-byte on external public sets (FLORES, BOUQuET, UD), lower is better.

Model	Params	FLORES	BOUQuET	UD
TATlit	478M	.740	.777	.502
Tweety-7B	7B	.757	.800	.544
Gemma-4-31B	31B	.772	.939	.691
Goldfish-tat	125M	.778	.801	.502
mGPT-1.3B	1.3B	.950	1.186	.880
gpt-oss-120b	120B	1.092	1.164	1.099
Qwen3-32B	32B	1.144	1.423	1.159
Llama-4-Scout	109B	1.222	1.498	1.080

Byte-per-byte on in-house held-out sets (literary, periodical), lower is better.

Model	Params	Held-out lit	Held-out per
TATlit	478M	.732	.701
Gemma-4-31B	31B	.790	.705
Tweety-7B	7B	.810	.778
Goldfish-tat	125M	.929	.874
Llama-4-Scout	109B	1.084	.995
Qwen3-32B	32B	1.105	.979
mGPT-1.3B	1.3B	1.199	1.094
gpt-oss-120b	120B	2.181	1.857

TatBLiMP, morphological acceptability, higher is better.

Model	Params	acc	acc_norm
TATlit	478M	.975	.958
Goldfish-tat	125M	.974	.958
Tweety-7B	7B	.956	.915
Gemma-4-31B	31B	.924	.839
Llama-4-Scout	109B	.889	.806
Qwen3-32B	32B	.811	.708
gpt-oss-120b	120B	.803	.677
mGPT-1.3B	1.3B	.736	.639

On the school-knowledge benchmark TUMLU-mini the model sits near chance (0.297 against a 0.25 random baseline), which is expected for a base model with no encyclopedic training.

Limitations

Base model, no instruction following, no encyclopedic knowledge. Softest axis is the person/possessive system inside izafet.

Authors

Ilshat Saetov, Dmitry Gaynullin.

Data

The training corpus is withheld on stewardship grounds for a minoritized language's literary heritage; a full datasheet with per-document provenance is available on reasonable request. The tokenizer and the model weights are released.

License

The model weights and the tokenizer are released under CC BY-NC-SA 4.0. TatBLiMP benchmark is released under CC BY-NC 4.0. The training corpus is not released.

Citation

bibtex

@software{saetov2026tatlit,
  author    = {Saetov, Ilshat and Gaynullin, Dmitry},
  title     = {TATlit --- a native literary Tatar base language model},
  year      = {2026},
  version   = {1.0},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.22212736},
  url       = {https://doi.org/10.5281/zenodo.22212736}
}