Patufet-educat

Crawls filtered for educational content

Overview 🌍

The patufet-educat dataset is a filtered version of the Catalan content in the Cultura X dataset, inspired by the fineweb-edu dataset, but specifically focused on educational content in Catalan.

All files 📂

The source code for creating this dataset can be found at the github repo, inside the educat folder.

You can view the prompt used here

Dataset filtering 🛠️

As explained on the fineweb-edu dataset card, filtering content for training datasets often involves using synthetic data to identify educational material. Here’s how we approached it:

Annotation ✏️

We scored 100,000 text samples from the Oscar corpus using the gemini-1.5-flash model. The prompt classified texts on a scale from 0 to 7, allowing for a finer distinction between content quality.

The annotated web samples can be found here.

Classification 🧠

We trained a text classifier using FastText on these annotations. Although a transformer-based model would likely improve dataset quality, resource constraints led us to use FastText. The classifier, available here, achieved a Precision and Recall of 0.51825 across nine categories (0-7, plus non-safe content).

Filtering ⚙️

We chose a threshold of 3, which is lower than the fineweb-edu threshold (our 3 is more permissive that the original one). This decision balances our classifier's performance with the limited availability of Catalan texts. Heavier filtering would have resulted in a much smaller dataset.

The classification and filtering process took 4:35 hours on a laptop. If you wish to use a different threshold, we’ve released the full tagged dataset.

In total, around 2/3 of the original dataset were filtered.

Problems Encountered 🚧

We opted for a 0 to 7 scale instead of 0 to 5 to better differentiate between low-quality and mediocre content. In our 100,000 annotated samples, there were very few high scores (none at 7, and only five at 6). To streamline the classifier, we consolidated these into a 5.

The Gemini API also flagged some samples as unsafe, excluding them from the dataset and marking them with a score of 999 (missing data). This filtering helps to remove potentially harmful content, although it may also exclude valuable educational material, such as topics related to sexual reproduction.

Evaluation 🧪

This dataset hasn't been evaluated because we don't have the resources to do so. However, viewing the results of fineweb-edu and other similar filtered datasets, we think that it's going to help improve the quality of Catalan LLMs

Considerations and Disclaimer ⚠️

While the patufet-educat dataset offers numerous benefits, there are some considerations to keep in mind:

• Remaining Harmful Content: Although we filtered out much harmful material, some undesirable content might still be present.
• Exclusion of Useful Content: Certain educational topics, such as those related to sexual reproduction, may have been unintentionally excluded due to the filtering process.
• Lack of Coding Content: Given the nature of Catalan content, there is likely a scarcity of programming or technical material in this dataset.
• Personal and Sensitive Information: As this dataset is derived from CommonCrawl, it may still contain personal or sensitive data. Users must consider this before using the dataset for tasks like training deep learning models.

License 📝

The licence should follow the terms for CulturaX, which in turn strictly follows those of mC4 and OSCAR. Please refer to both below licenses when using this dataset.

• mC4 license
• OSCAR license

Conclusion and Recommendations

The creation of the patufet-educat dataset highlights both the challenges and opportunities involved in curating educational content in Catalan. While we were able to filter a significant amount of data, several key issues emerged that are important for anyone looking to replicate or build upon this work.

• Limited Availability of Catalan Content: We detected a limited availability of high-quality educational content in catalan, which made us select a lower threshold for filtering. This also lead to the creation of other datasets, like patufet-textbooks
• Quality vs Quality: Given the smaller pool of Catalan content, we faced a trade-off between quality and quantity. To ensure a sufficiently large dataset, we included some "meh" quality content, which might not be ideal for all use cases.
• Sensitive topics: Some sensitive topics that are important and not harmful have been left out due to the high safety settings in the Gemini API
• Improved Filtering Models: If resources allow, using a transformer-based model for classification would likely yield a higher-quality dataset.
• Broader Data Collection: o address the scarcity of Catalan content, future efforts could focus on broader data collection, perhaps even scraping specific educational websites or creating more targeted crawls that are likely to contain valuable material.

Despite these challenges, the patufet-educat dataset represents a valuable resource for developing educational tools and language models in Catalan. While there is room for improvement, this dataset provides a strong foundation that can be built upon in future iterations. By sharing the process and considerations, we hope to facilitate further development in this area and encourage collaboration among researchers and developers working with low/medium-resource languages :).

---

For more detailed information on the process of creating this dataset or any other inquiries, feel free to reach out. 😊