Datasets:
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.
Til-Qazyna kk corpus. Request access for research/non-commercial use.
Log in or Sign Up to review the conditions and access this dataset content.
Til-Corpus exp078 (grounded, kk-dominant)
Kazakh-dominant pretraining corpus, curriculum-ordered (quality rises toward the end of training).
Version exp078-grounded-v1 (fingerprint eb97c56272582f74). Parent: exp072-regrade-base.
Composed (training-ready) — composed/
- 7.21B tokens / 32.9M rows (Til-Tokenizer-128k).
bulk.jsonl— q4 tier, 4.83B tok / 15.69M rows. Read FIRST.anneal.jsonl— q5 (premium) tier, 2.38B tok / 17.23M rows. Read LAST (LR-decay imprints quality).- Trainer MUST read
bulk.jsonlTHENanneal.jsonl, no global shuffle, so the WSD decay phase (~last 30%) trains on q5. q5 share = 33% tok / 52% rows (>=30%). dataset_manifest.json,compose_stats.json,CHANGELOG.md.
Components (for re-compose / iteration)
final_ground{,2,3,4}/final.jsonl— 4 grounded anti-calque batches over native kk passages 0–7.25M: 13.7M examples / 1.47B tok. Generated FROM real kk text (no calque), gen = Qwen3-VL-30B + GPT-OSS-120B, judge = single-Qwen q>=4. Tasks: summ/qa/explain/test/keypoints/expand + title_kw/classify/qa1.final/kept_*.jsonl— regrade base (~5.74B):kept_curated(kk native passages) +kept_gen73+kept_ru+kept_en+kept_code+kept_math, all Qwen-regraded q>=4 with category tags.
Iteration cycle
Add a new grounded batch -> re-run gen78_compose.py -> gen78_manifest.py bumps the version +
fingerprint -> new version is diffable against exp078-grounded-v1.
- Downloads last month
- 9