KazBench — Kazakh Language LLM Benchmark

KazBench is an open benchmark for evaluating large language models on the Kazakh language (a low-resource language). This repository hosts the public DEV split. A private TEST split is held out for anti-contamination and is used to verify leaderboard submissions.

• Code & harness: https://github.com/Yersultan04/kazbench
• Leaderboard (HF Space): https://huggingface.co/spaces/Yersultan03/kazbench-leaderboard
• License: CC-BY-4.0 (data) · MIT (code)

Tasks (6)

Config	Task	Metric
knowledge_mc	World/Kazakhstan knowledge, multiple choice	accuracy
reading_comprehension	Reading comprehension MC	accuracy
grammar_morphology	Grammar & morphology MC	accuracy
sentiment	Sentiment classification	accuracy
translation	RU/EN → KK translation	chrF
instruction_following	Instruction following (open-ended)	LLM-judge

Splits

This repo = public DEV (600 items, 100 per task). Of these, 296 are native-validated (validated: true) and 304 are pending native review (validated: false). The pending pool now includes real human-sourced data integrated from open datasets, not only seed items — see the source field:

• source: exam — real ЕНТ/ҰБТ exam questions (from kz-transformers/kazakh-unified-national-testing-mc, Apache-2.0) in knowledge_mc
• source: community — human-labeled reviews (from Darmm/darmm-sentiment-kk, Apache-2.0) in sentiment
• source: seed — model-generated items pending native review

Filter on validated for the headline-quality subset, or on source by provenance.

A private TEST split (180 items) is held out to prevent training-set contamination; submit predictions via PR to the GitHub repo and they are scored against TEST.

Item schema

Each line is a JSON object. Multiple-choice tasks:

{"id": "kmc_000001", "task": "knowledge_mc", "source": "seed",
 "validated": true, "question": "...", "choices": ["...","...","...","..."],
 "answer": 0}

See schema.md for the full data contract (all task types, canary items, results format).

Provenance & validation

Items are seed-generated and native-validated by a Kazakh speaker (validated: true). The benchmark includes canary items for contamination detection.

Citation

@misc{kazbench2026,
  title  = {KazBench: An Open Benchmark for Evaluating LLMs on the Kazakh Language},
  author = {Yersultan},
  year   = {2026},
  url    = {https://github.com/Yersultan04/kazbench}
}