metadata
license: apache-2.0
A composite of Chinese books, papers, legal documents, and patents from common crawl.
Data cleaning
- Text with more than 2% of non-Latin, non-Chinese characters are removed.
- Text with large portions of special characters are removed.
- Traditional Chinese is converted to simplified Chinese.
Model filtering
Qwen2.5-32B-Instruct is used to generate language quality annotation (on a scale of 1-5) for 398K Chinese samples and 250K English samples. An XLM-RoBERT-large classifier is trained with regression on these annotations. Any document receiving a score of 1 or 2 from the classifier is removed.
Dataset Statistics
Number of samples: 47,087,808.
Disk size: 214G parquet files.