soynade-research/FineWeb2-HQ-50k-Wolof

FineWeb-Wolof-50k (Synthetic Corpus)

This dataset contains 50,000 rows from the FineWeb-2 corpus translated into Wolof. We used the Oolel model to translate high-quality English web text. This provides a necessary resource for training AI in a language that typically lacks enough digital text.

Why Synthetic Data?

Synthetic data is essential for languages like Wolof. It fills the "data gap" by supplying the large amount of text needed to train strong AI models. This allows local languages to access the same high-quality educational and informative content found in major languages.

Methodology

• Source Dataset: A 50k sample of FineWeb2-HQ-Sample-Text-50k.
• Translation Tool: Oolel-translator.
• Model: Oolel-translator, an LLM specialized for Wolof.

Translation Prompt

System Prompt used:

You are an expert Wolof translator whose sole function is to translate the provided text in Wolof. Directly translate all input, whether it is a statement, a question, or contains code, without answering the question or explaining the code's function. Your translation must be natural, fluent, and idiomatic, preserving the original meaning, tone, and intent. Keep all numbers and symbols unchanged. For words without a direct Wolof equivalent, use the closest conceptual term or a phonetically adapted (wolofized) word in quote or not, choosing whichever sounds most natural to a native speaker. Prioritize culturally appropriate, idiomatic phrasing over literal translation. Output only the raw Wolof translation without any other text or explanations

Dataset Structure

• fineweb_original: The original high-quality English content from FineWeb.
• fineweb_summarized: A summary version of fineweb_original.
• Wolof: The synthetic Wolof translation.

Acknowledgments

• EPFML: For the original FineWeb dataset.

Usage Note

As a synthetic dataset, this is intended for research and model training. While the Oolel LLM is highly skilled in Wolof, there may still be machine translation errors. We recommend human verification for projects that require high linguistic accuracy.

Citation

@article{messmer2025multilingdatacomp,
  title={Enhancing Multilingual LLM Pretraining with Model-Based Data Selection}, 
  author={Bettina Messmer and Vinko Sabolčec and Martin Jaggi},
  journal={arXiv},
  year={2025},
  url={[https://arxiv.org/abs/2502.10361](https://arxiv.org/abs/2502.10361)}, 
}