FineWeb2 Edu Japanese: 120 Million Educational Texts with 89B Tokens

Encompassing 120 million Japanese educational texts totaling 89.3 billion tokens, released by hotchpotch in May 2025. It represents a high-quality subset filtered from the 376 million Japanese records in the FineWeb2 corpus based on educational value.

Use Cases

• Pre-training Japanese large language models using the 89.3B token corpus
• Fine-tuning models on short-form educational content using the 512-token subset
• Benchmarking Japanese text quality classifiers against the 120M educational records

Strengths

• 120 million filtered records
• 89.3 billion tokens for large-scale training
• Includes a 10B token sample for lower-compute experimentation
• Specific educational quality filtering applied to raw web crawl

Limitations

• Potential for algorithmic bias inherent in the automated educational filtering process
• Lack of manual human verification for the 120 million individual records
• Geographic bias toward web-accessible Japanese content

Provenance

Source

FineWeb2

Collection Method

Algorithmic filtering of 376 million Japanese web-scraped texts

Freshness

Last updated May 2025; derived from FineWeb2 web-scraped data.

Geography

Japan