Til-Zharar-v2

Kazakh essays, compositions (қмж), term papers (реферат) and poems (өлең) from zharar.com, with full document text recovered from the site's downloadable files. Part of TilQazyna.

TL;DR

• processed = judged/tiered (training); raw = pre-judge.
• v2 over v1: we download every attached file (docx/pptx/pdf) and extract its full text (paragraphs + tables), which is often richer than the on-page preview (e.g. a қмж page shows ~63k chars but its file holds ~140k). Page text is the fallback.
• Judged with Qwen → {score, category, lang}, kk_ratio, tier.

Schema (processed)

url, category, slug, dl_id, title, page_text, doc_text, text, ext, n_images, is_scan, score, category(judge), judge_lang, kk_ratio, tier, source, n_chars (text = the richer of file vs page; is_scan flags documents whose content is embedded images.)

Usage

from datasets import load_dataset
ds = load_dataset("TilQazyna/Til-Zharar-v2", "processed", split="train")

Limitations & ethics

User-generated study material — quality varies (filter by tier). Some documents are image scans (is_scan=true) whose pictures are not yet OCR'd. Non-commercial Kazakh-language research use.