SozKZ Corpora: Kazakh Training Datasets
Collection
Training corpora for Kazakh LLMs — raw, cleaned, deduplicated, tokenized, synthetic, and parallel datasets • 22 items • Updated
Datasets:
Kazakh Clean Pretrain v2 (Tokenized)
Pre-tokenized Kazakh corpus for LLM training. Each sample is a packed block of 1024 tokens.
| Property | Value |
|---|---|
| Train blocks | 1,006,344 |
| Val blocks | 10,059 |
| Block size | 1024 tokens |
| Total tokens | ~1.04B |
| Tokenizer | stukenov/kazakh-gpt2-50k |
| Vocab size | 50,257 |
Domain distribution
| Domain | Documents | Share |
|---|---|---|
| cc100 | 16,408,890 | 87.4% |
| leipzig | 1,615,760 | 8.6% |
| kazakhNews | 302,889 | 1.6% |
| wikipedia | 231,654 | 1.2% |
| oscar | 204,826 | 1.1% |
| kazakhBooks | 4,044 | 0.0% |
- Downloads last month
- 10