Geralt-Targaryen/CC-zh

High-quality Chinese text from Common Crawl cleaned by the following steps:

• Documents containing more than 2% non-Chinese, non-English characters are removed. Those containing more than 30% digits or capital letters are also removed.
• Documents whose language is identified as non-Chinese by fasttext are removed.
• All text in Traditional Chinese is converted into Simplified Chinese.
• Low-quality documents (e.g. boilerplates, advertisements) are heuristically removed based on statistics such as average line length, portion of special characters, etc.
• Exact deduplication is performed in buckets of around 100GB compressed text. We did not deduplicate globally due to memory constraints, and estimate that about 0.03% of the documents are exact duplicates based on small-scale cross-bucket deduplication.
• Qwen2.5-32B-Instruct is used to generate language quality annotation (on a scale of 1-5) for 9.3M Chinese documents and 9.2M English documents, from which we sample 398K Chinese documents and 250K English documents to balance label distribution. An XLM-RoBERT-large classifier is trained with regression on these annotations. Any document receiving a score lower than 4 is removed.

Details about Model Annotations

On 2K samples, we compared the annotation distribution (in percentage) of Qwen2.5-Instruct 32B and 72B:

Score	1	2	3	4	5
32B	0.7	17.1	45.7	35.8	0.8
72B	0.3	4.7	22.9	58.1	14.1

The scores between the two models have a correlation coefficient of 0.75, and manual inspection suggests that both are satisfactory. We eventually choose the 32B model for both efficiency and more balanced label distribution.