A unified pretraining-text addition for the Kazakh Til-Corpus: the judged, high-quality text gathered from TilQazyna's 2026 new sources — KazNEB OCR books, zharar essays, the KK educational web crawl, and 600-kitap audiobook ASR transcripts. Realizes the master-plan rule "recognition/scrape text → ALSO Til-Corpus".

What's in it

Only premium + clean rows (judge score ≥3) from each source's processed/ split are kept, unified to a single text schema. Each row carries its provenance.

field	meaning
`text`	the document/page/transcript text
`lang` / `score` / `category`	judge labels (Qwen)
`tier`	`premium` (score≥4 & Kazakh) / `clean`
`source`	Til-Books-KazNEB / Til-Zharar-v2 / Til-Web-KK / Til-Audio
`route`	`ocr` / `web` / `asr`

Usage

from datasets import load_dataset
ds = load_dataset("TilQazyna/Til-Corpus-Additions-v2", split="train")
premium = ds.filter(lambda r: r["tier"] == "premium")

Intended to be deduped + merged into a future Til-Corpus v2. Non-commercial Kazakh-language research.

Downloads last month: 262

Total file size:

6.09 GB