Til-Web-KK

Clean Kazakh web text extracted from a crawl of Kazakh educational / reference sites (okulyk.kz, mektep.kz, emektep.kz, referat.resurs.kz, surak.baribar.kz, ref-kaz999, almatykitap.kz, expert.study-inn.kz, mhelp.kz). Part of the TilQazyna corpus program.

TL;DR

• processed config = judged, tiered text (use for training); raw = pre-judge extraction.
• One row per page: main-content text (boilerplate removed) + quality score/category/lang/tier.
• High quality: ~62 % of pages are premium (score≥4) — far cleaner than forum/essay scrapes, because these are curated educational pages.

Pipeline

1. Crawl each site (staged, resumable; pages stored as pages.tar.zst).
2. Extract main content with trafilatura (navigation/ads/boilerplate stripped; corrupt archives and pathological HTML are skipped with a per-page timeout so the run never stalls).
3. Judge with a Qwen judge → {score 1–5, category, lang}, kk_ratio, and tier (premium = score≥4 & Kazakh; clean = score≥3; else raw).

Schema (processed)

url, site, path, title, text, lang, score, category, judge_lang, kk_ratio, tier, source, n_chars

Usage

from datasets import load_dataset
ds = load_dataset("TilQazyna/Til-Web-KK", "processed", split="train")
prem = ds.filter(lambda r: r["tier"] == "premium")

Limitations & ethics

Web text may include duplication and site furniture residue; filter by tier/kk_ratio. Collected for non-commercial Kazakh-language research; respect the source sites' rights.