A large-scale cleaned and deduplicated Kazakh text corpus assembled from 18 public sources. Designed for pre-training causal language models on Kazakh text.

	Overview

Total texts 13,700,018

Train split ~13,563,018 (99%)

Validation split ~137,000 (1%)

Raw input 28,431,116 texts

Pass rate 48.2%

Dedup removed 3,170,330

License Apache 2.0

	Sources

Source Raw Clean Pass… See the full description on the dataset page: https://huggingface.co/datasets/stukenov/sozkz-corpus-clean-v3.