TatBLiMP — Tatar Benchmark of Linguistic Minimal Pairs

⚠️ Work in progress (v0, 2026-07). Current build: 13 of 16 target phenomena, 1010 ratified pairs (md5 788e14…). Data files will be uploaded with the public release.

Summary

TatBLiMP is a minimal-pairs benchmark for Tatar (tt, Cyrillic) morphosyntax, in the tradition of BLiMP (Warstadt et al. 2020) and adapting most of TurBLiMP's phenomena to Tatar. Each item is a pair of sentences differing by one minimal edit — one grammatical (good), one ungrammatical (bad). A model passes an item when logprob(good) > logprob(bad). No generation, no parser — runs on base models and mid-training checkpoints.

Method

good sentences come from three sources: hand-adapted TurBLiMP items (≈38%; government and lexicon adjusted to Tatar norms, Turkish–Tatar false friends kept as targeted hard items), attested sentences from a curated pool of Tatar prose and poetry (≈34%), and hand-composed items built from corpus-verified vocabulary and collocations (≈28%). bad is a clone with exactly one edit — a morpheme swap built with the apertium-tat transducer, a licensed insertion, or a word-order permutation. Every pair is ratified by a native speaker.

Coverage (v0)

phenomenon	pairs
anaphor agreement	80
argument structure — ditransitive	78
argument structure — transitive	78
binding (Principle B)	78
ellipsis (backward gapping)	85
irregular forms	80
island effects (wh-word at clause boundary)	80
nominalization	73
NPI licensing	80
passives (case of the passive subject)	83
relative clauses	63
subject–verb agreement	77
word order (scrambling in embedded clause)	75

Remaining phenomena are in progress.

Format: JSONL — good, bad, phenomenon, swap, src, book_id, kind. Metric: acc_norm (byte-normalized) primary, acc secondary.

License

TBD — will be finalized before the data release.

Citation

@misc{tatblimp2026,
  title  = {TatBLiMP: A Benchmark of Linguistic Minimal Pairs for Tatar},
  author = {Ilshat Saetov},
  year   = {2026},
  url    = {https://huggingface.co/datasets/ilchats/TatBLiMP}
}

Built on BLiMP (Warstadt et al. 2020), TurBLiMP (arXiv:2506.13487) and the apertium-tat transducer (Washington, Salimzyanov & Tyers, LREC 2014).