---
license: odc-by
language:
- en
---
Dataset Summary

![image/png](https://cdn-uploads.huggingface.co/production/uploads/6328952f798f8d122ce62a44/0gun4rh6OgCgew5J6HMv5.png)

This dataset is a filtered subset of the Fineweb-Edu-Fortified dataset. The primary goal of this subset is to reduce the dataset size to a more manageable volume while maintaining high-quality content. It contains three key fields: score, text, and url, focusing on entries with a score of 4 and above, indicating higher relevance and quality of educational content.

This dataset can be used for several fine-tuning and model improvement tasks, including model healing, synthetic dataset creation, and topic-specific fine-tuning for smaller language models.
Source and Reference

The original dataset, Fineweb-Edu-Fortified, is a deduplicated version of the Fineweb-Edu dataset. Fineweb-Edu was created from data processed across 95 crawls of Common Crawl, covering web content from 2013 to 2024. This filtered version maintains the core educational texts while stripping out unnecessary fields to make the dataset more accessible for specific machine learning tasks.

The original dataset is detailed in its dataset card here.
https://huggingface.co/datasets/airtrain-ai/fineweb-edu-fortified

Score: Represents the quality of the text content.
Text: The main content of the webpage 
URL: The source URL from which the text was extracted.

Use Cases

This filtered version has several potential use cases, including:

Model Healing: Can be used as fine-tuning data for restoring or enhancing model knowledge after Depth Up scaling or model pruning. The high-quality, curated educational content ensures that pruned models can recover key domain knowledge effectively.

Synthetic Dataset Creation: The dataset can be used in a pipeline to generate question-answer pairs by extracting sections of the text, generating relevant questions, and providing the answers. This makes it useful for data augmentation and Q&A models.

Smaller Dataset Training: It can be employed for training smaller LLMs by utilizing the filtered high-quality educational text, especially useful in cases where resource constraints exist.

Topic-Based Fine-Tuning: The dataset can be grouped by subject or topic based on the content of the text or URL, allowing for the creation of more specialized training datasets tailored to particular domains or educational levels.

License

This dataset is released under the Open Data Commons Attribution License (ODC-By) v1.0, in accordance with the licensing of the original dataset.
Citation

If you use this dataset, please cite the original dataset: https://huggingface.co/datasets/airtrain-ai/fineweb-edu-fortified



@dataset{airtrain2024finewebedu,
  title={Fineweb-Edu-Fortified},
  author={Airtrain AI},
  year={2024},
  url={https://huggingface.co/datasets/airtrain-ai/fineweb-edu-fortified}
}