MIyamoto Keisuke
Update README.md
8b2552e verified
|
Raw
History Blame Contribute Delete
1.8 kB
metadata
license: odc-by
language:
  - ja
size_categories:
  - 10M<n<100M
task_categories:
  - text-generation
pretty_name: CleanedFineWeb2Edu-jp

CleanedFineWeb2Edu-jp

CleanedFineWeb2Edu-jp is a cleaned Japanese web text dataset.

This dataset was created from the sample_10BT subset of hotchpotch/fineweb-2-edu-japanese. The source text was refined with MK0727/corpus-refiner-jp.

Purpose

The main purpose of this dataset is to provide cleaner Japanese web text for language model pretraining and continued pretraining.

This dataset keeps Japanese web documents from FineWeb2-Edu while reducing boilerplate, navigation fragments, repeated links, metadata-like text, and other low-value lines. It is intended for use cases that need broad Japanese web text with less page-structure noise than the original subset.

Dataset Creation

After applying MK0727/corpus-refiner-jp, documents were filtered out when the refined text was 200 characters or shorter, or when it was 4096 tokens or longer with sbintuitions/modernbert-ja-130m.

Dataset Statistics

Item Value
Source size about 12.8B Gemma 4 tokens
Final size about 9.95B Gemma 4 tokens
Retention rate about 78%

Data Fields

Field Description
text Cleaned Japanese text
id Original document ID
dump Common Crawl dump name
url Source URL
date Crawl date
file_path Original Common Crawl file path
output_char Character count of text
output_token Token count of text using google/gemma-4-26B-A4B-it

License

This dataset is released under the ODC Attribution License (odc-by).