Survivor: 350B+ Educational Tokens from Multilingual PDFs

FinePDFs-Edu dataset consists of over 350 billion tokens of educational text extracted from PDFs. The dataset was created by Web3Survivor using an educational quality classifier trained on annotations from a large language model to filter content from the broader FinePDFs collection. It covers 69 languages and was last updated on December 11, -0001.

Use Cases

• Training multilingual language models based on the 350B+ token scale of educational text.
• Fine-tuning models for educational content generation based on the PDF-derived text.
• Conducting comparative linguistic analysis based on the dataset's coverage of 69 languages.
• Developing domain-specific classifiers based on the educational quality filtering methodology.

Strengths

• Contains over 350 billion tokens of text, indicating a large scale.
• Covers educational content across 69 languages.
• Filtered using a dedicated educational quality classifier to improve content relevance.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count and file formats are unknown, which may limit suitability assessment.
• Data may reflect source bias inherent to the PDFs used for extraction.

Provenance

Source

Web3Survivor, via Hugging Face.

Collection Method

Extracted and filtered from the FinePDFs dataset using an educational quality classifier trained with annotations from a Qwen3-235B-A22B-Instruct-2507 model.

Time Range

null

Freshness

Last updated 2025-12-11 09:36:49; freshness should be verified.

Geography

null