FineWeb-Edu Score 2: 5.4 Trillion Tokens of Educational Web Content

5.4 trillion tokens of educational web content filtered from the FineWeb corpus by HuggingFaceFW, updated in 2025. It captures a broad range of instructional material by applying an educational quality classifier with a threshold score of 2.

Use Cases

• Pre-training large language models using the 5.4 trillion tokens of text content
• Evaluating educational quality distributions using the score 2 threshold metadata
• Fine-tuning models on broad instructional web data

Strengths

• 5.4 trillion tokens
• ODC-By license
• Filtered using an educational quality classifier

Limitations

• Lower quality threshold (score 2) results in higher noise levels than the 1.3T token version
• English-only language focus as indicated by tags

Provenance

Source

HuggingFaceFW (FineWeb project)

Collection Method

Web scraping followed by automated classifier filtering

Freshness

Last updated July 2025

Geography

Global (web-based)