You need to agree to share your contact information to access this dataset
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
This dataset is a raw mirror of third-party Kazakh educational websites, collected for non-commercial NLP research (the Til foundation-model program). Access is granted manually. By requesting access you agree to use the data for research purposes only and to respect the original sites' terms.
Log in or Sign Up to review the conditions and access this dataset content.
Til Web Raw Corpus — KK (v1)
Raw web crawl of Kazakh educational / Q&A websites, collected for the Til Kazakh foundation-model corpus program. This is a raw mirror (full server HTML + in-content images and downloadable attachments), not a cleaned text corpus — text extraction / cleaning is a downstream step.
Collection respected each site's robots.txt (only sitemap-listed content URLs
were fetched; admin/login/create paths excluded). Crawl date: 2026-06-21/22.
Sites in this release
| Site | Pages | Raw HTML | Assets (img+files) | Notes |
|---|---|---|---|---|
bilim-all.kz |
94,777 | 8.2 GB | 25,323 (25,312 img + 70 docs) | articles, өлең (poems), quotes, lesson plans (jospar), puzzles, names |
surak.baribar.kz |
128,113 | 9.2 GB | 35 | Question2Answer Q&A; all answers server-rendered & captured (verified saved==live) |
Pages with 404 (deleted articles): 10 on bilim-all. surak ~10% of questions are genuinely unanswered (no answers exist on the live site).
Layout
<host>/
pages.tar.zst tar of the HTML mirror (extract: zstd -d | tar -x)
assets.tar.zst tar of downloaded images + attachments
urls.txt full list of crawled URLs (from sitemap)
pages_manifest.jsonl.zst one line per page {url,status,path,bytes,ts}
assets_manifest.jsonl one line per downloaded asset
asset_urls.txt discovered asset URLs (kind<TAB>url)
Inside pages.tar.zst, files mirror the URL path
(article/5.html, 100030/index.html, ...). To map an HTML file back to its
URL, use pages_manifest.jsonl.
Method
Collected with a staged, resumable crawler
(scripts/crawl/site_crawler.py in the Til repo): build-urls (expand sitemap
index) → fetch-pages (mirror HTML) → discover (parse content assets) →
fetch-assets (download images + linked documents). Polite concurrency, no
rate-limit (HTTP 429) hits.
Intended use
Pretraining-corpus source material for Kazakh LLMs. Raw HTML — run boilerplate removal / text extraction before training. Non-commercial research use; respect the source sites' rights.
- Downloads last month
- 11