AINA-KK

Benchmark for detecting machine-generated Kazakh text, from the paper "Unchecked or Falsely Accused? Evaluating AI-Generated Text Detectors on Kazakh" (Mansur Ozaman, Nazarbayev University), arXiv:2610.15870. Code: https://github.com/ozxmn/aina-kk. Detector: https://huggingface.co/ajnasiou/aina-kk-det.

Contents

File	Rows	What it holds
data/human.jsonl	3,811	Human passages written before 30 Nov 2022: encyclopedic (Kazakh Wikipedia dump of 2022-05-01, with text), news, web (CC100-kk) and educational pages (identifiers and URLs only).
data/machine.jsonl	5,597	Machine text (M), human-opening + machine-continuation passages (MIX), AI-polished human text (HP), machine translation from Russian (MT) and humanizer rewrites (MH), from 9 generators.
data/eval_index.jsonl	14,583	Every row of the evaluation file, including the stress-test variants (Latin-2021 transliteration and its inverse, homoglyphs, truncation, folded homoglyphs, truncated development passages). Text is omitted for rows derived from non-Wikipedia human passages.
scores/		Per-passage scores of every detector in the paper: fine-tuned (finetuned_A; finetuned_K for the runs with KazLLM and Sherkala in training), zero-shot, public detectors. finetuned_B holds leave-one-generator-out runs and runs trained on all eight generators, which the paper does not report.
scores/release_model/		Scores of the released AINA-KK-Det model: its three classifiers on the benchmark file, and the full detector (text preparation, classifier and Binoculars score) on every non-training row.
audit/	100 + 273 + 100 + 100	Validity judgments: the first native-speaker audit (audit1_human.jsonl) and the LLM screen (Claude Opus 5.5) of the same pool (audit1_llm_screen.jsonl), both before the last cleaning pass; a fixed-prompt re-run of that screen on the 100 annotated items, after the benchmark was frozen (llm_screen.jsonl), and a native-speaker audit of a fresh random sample of the final benchmark (audit2_postfilter.jsonl: 13 human passages per domain, 6 machine passages per generator).

Fields: split (the passage's role: train or test; every held-out-generator row and every MIX, HP, MT and MH row is marked test), src_split (the topic split of the source passage; use it for topic-disjoint analyses), cluster (topic cluster; the 7 development clusters, held out of classifier training and used for thresholds, are 3, 7, 29, 30, 42, 55 and 59), domain, label (H, M, MIX, HP, MT, MH), generator, condition (C1 topic-to-text, C2 continuation, C3 polishing, C4/C5 translation, C6 humanizing), heldout (generator never used for training), group/source_passage (the human passage a machine text was prompted from), variant.

Human text

Wikipedia passages are included under CC BY-SA 3.0, the license of the 2022 dump. News and educational passages are given as their URL and FineWeb2 record identifier (HuggingFaceFW/fineweb-2, kaz_Cyrl); web passages come from CC100-kk. Two kinds of machine row carry human text from these sources: mixed-authorship (MIX) rows begin with the first ~50 words of their source passage, and AI-polished (HP) rows are model rewrites of a whole source passage. The code repository contains the extraction and cleaning steps used to build them from the public source files. Emails, phone numbers and 12-digit ID numbers were replaced by the placeholders [EMAIL], [PHONE] and [IIN]; machine text never contains them, so remove or neutralise them before training a detector.

Data quality

In the final audit, 33 of the 52 sampled human passages were judged valid (66% with domains weighted by size, 95% CI 52–79%). The rest were judged broken or truncated (12), to read like machine translation (5), boilerplate (1) or partly Russian (1). A date before ChatGPT rules out modern chatbots, not earlier machine translation. The benchmark was frozen before this audit and is released as it was evaluated; use the audit identifiers to study or exclude these passages.

Models used and their terms

Role	Model	License or terms
Generator (API)	Gemini 3.1 Flash-Lite, Gemini 3.5 Flash-Lite, Gemma 4 31B, Gemma 4 26B-A4B (Gemini API)	Gemini API terms; Gemma 4 weights Apache 2.0
Generator (API)	gpt-oss-120b, Qwen3.8-27B (Groq)	Apache 2.0 weights; Groq terms
Generator (open)	issai/LLama-3.1-KazLLM-1.0-8B	CC BY-NC 4.0
Generator (open)	inception42/Llama-3.1-Sherkala-8B-Chat	CC BY-NC-SA 4.0; its card also requires adherence to the Llama 3.1 Community License
Translation	facebook/nllb-200-distilled-600M	CC BY-NC 4.0
Zero-shot scorers	google/gemma-4-E2B(-it), Qwen/Qwen3-4B(-Base), unsloth/Llama-3.2-3B(-Instruct), ai-forever/mGPT, ai-forever/mGPT-1.3B-kazakh	Apache 2.0; Apache 2.0; Llama 3.2 Community License; Apache 2.0; MIT
Fine-tuned encoders	microsoft/mdeberta-v3-base, FacebookAI/xlm-roberta-base	MIT
Public detectors	openai-community/roberta-base-openai-detector, desklib/ai-text-detector-v1.01, Oxidane/tmr-ai-text-detector, fakespot-ai/roberta-base-ai-text-detection-v1, Hello-SimpleAI/chatgpt-detector-roberta	MIT; MIT; MIT; Apache 2.0; no license stated
Language identification	cis-lmu/glotlid	Apache 2.0 with notices
Sources	Kazakh Wikipedia (2022-05-01 dump), FineWeb2 kaz_Cyrl, CC100-kk	CC BY-SA 3.0; ODC-By 1.0; Common Crawl terms

Licenses as listed on each model's Hugging Face page on 11 October 2026.

Intended use

Research on detecting machine-generated text and on the false-positive behaviour of detectors. The data must not be used to build tools that judge individual writers without human review: no detector evaluated in the paper should serve as sole evidence of misconduct.

Removal requests

If a passage in AINA-KK is yours and you want it removed, write to mansur.ozaman@nu.edu.kz with its identifier or URL. It will be removed in the next release.

License

Machine-generated text, splits, annotations and scores: CC BY-NC 4.0. Wikipedia text: CC BY-SA 3.0. Generated text was produced with the models listed in the paper and remains subject to their providers' terms.

Citation

@misc{ozaman2026aina,
  title         = {Unchecked or Falsely Accused? Evaluating AI-Generated Text Detectors on Kazakh},
  author        = {Ozaman, Mansur},
  year          = {2026},
  eprint        = {2610.15870},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CL},
  url           = {https://arxiv.org/abs/2610.15870}
}