Yxanul/fineweb-edu-highest-quality-2025

FineWeb-Edu Highest Quality Dataset (2025 Collection)

Dataset Summary

This dataset contains 4.17 billion tokens of the highest quality educational content, carefully filtered from the FineWeb-Edu dataset's 2025 Common Crawl snapshots. This represents the cream of the crop - only the top ~2% of documents that meet strict quality criteria.

Key Statistics

• Total Tokens: 4,176,738,951
• Total Documents: 1,477,151
• Average Tokens per Document: 2,827
• Storage Size: ~11 GB (compressed parquet)
• Source Snapshots: 6 CC-MAIN-2025 snapshots

Quality Filtering Criteria

Each document in this dataset meets ALL of the following strict criteria:

1. Token Length: ≥ 1,000 tokens (ensures substantial, detailed content)
2. Educational Score: ≥ 3.5 out of 5 (top 15% educational quality)
3. Language Score: ≥ 0.95 confidence (ensures high-quality English)

Filtering Impact

Metric	Original	Filtered	Retention Rate
Documents	~84M	1.47M	1.75%
Tokens	~108B	4.17B	3.86%

This aggressive filtering ensures every document is:

• Educationally valuable
• Well-written and coherent
• Substantial in length
• Grammatically correct

Source Data

Data sourced from HuggingFaceFW/fineweb-edu, specifically these 2025 snapshots:

Snapshot	Original Rows	Filtered Tokens
CC-MAIN-2025-26	14M	589M
CC-MAIN-2025-21	15.6M	661M
CC-MAIN-2025-18	17.3M	731M
CC-MAIN-2025-13	16.8M	702M
CC-MAIN-2025-08	16.1M	678M
CC-MAIN-2025-05	19.3M	815M

Dataset Structure

Fields

• text (string): The main text content
• id (string): Unique identifier
• dump (string): Source CC-MAIN snapshot
• url (string): Original URL
• date (string): Crawl date
• file_path (string): Original S3 path
• language (string): Detected language (always 'en')
• language_score (float): Language detection confidence (≥0.95)
• token_count (int): Number of tokens (≥1000)
• score (float): Educational quality score (≥3.5)
• int_score (int): Integer educational score (4 or 5)
• source_snapshot (string): CC-MAIN snapshot identifier

File Format

• Format: Parquet with Snappy compression
• Files: Multiple parquet files for easy streaming
• Naming: {snapshot}_{batch_number}.parquet

Intended Use

This dataset is ideal for:

• Pre-training smaller language models (1B-7B parameters)
• Fine-tuning existing models for educational tasks
• Continued pre-training to improve educational capabilities
• Research on high-quality educational text

Quality Examples

The filtering process ensures content like:

• Detailed technical tutorials
• Comprehensive educational articles
• Well-structured explanations
• Academic discussions
• Professional documentation

Processing Pipeline

1. Download: Raw CC-MAIN snapshots from FineWeb-Edu
2. Filter: Apply strict quality criteria (score ≥3.5, tokens ≥1000, language ≥0.95)
3. Compress: Save as Snappy-compressed Parquet
4. Validate: Verify all quality thresholds maintained
5. Merge: Combine batches with source tracking

Comparison with Other Datasets

Dataset	Tokens	Quality Filter	Avg Doc Length
This Dataset	4.17B	Very High (top 2%)	2,827 tokens
The Pile	825B	Mixed	Variable
RedPajama	1.2T	Medium	~2,000 tokens
FineWeb-Edu (full)	1.3T	Medium	~1,100 tokens

Citation

If you use this dataset, please cite:

@misc{fineweb-edu-highest-quality-2025,
  title={FineWeb-Edu Highest Quality Dataset (2025 Collection)},
  author={Yxanul},
  year={2025},
  publisher={HuggingFace},
  howpublished={\url{https://huggingface.co/datasets/Yxanul/fineweb-edu-highest-quality-2025}}
}

Also cite the original FineWeb-Edu:

@software{penedo2024fineweb-edu,
  author = {Penedo, Guilherme and Kydlíček, Hynek and Cappelli, Alessandro and Sasko, Mario and Wolf, Thomas},
  title = {FineWeb-Edu: A openly accessible multilingual dataset for educational content},
  month = May,
  year = 2024,
  url = {https://huggingface.co/datasets/HuggingFaceFW/fineweb-edu}
}

License

This dataset inherits the license from FineWeb-Edu. Please refer to the original dataset for licensing information.

Acknowledgments

• Thanks to HuggingFace FineWeb team for the original dataset
• Filtered using custom scripts optimized for highest educational quality
• Processing supported by local compute resources

Contact

For questions or issues, please open a discussion on the HuggingFace dataset page.