Tropic-AI/moberto-pretraining-dataset-c4-compatible

moBERTo Pretraining Dataset

This dataset is the curated Portuguese corpus used for the continued pretraining of moBERTo, a Portuguese adaptation of ModernBERT. It contains approximately 12 billion tokens of Portuguese web text, filtered for educational and STEM content. The data is stored in a C4-compatible format to simplify integration with standard pretraining pipelines.

Dataset Summary

The corpus is built from the Portuguese subset of FineWeb2, a large-scale web dataset derived from CommonCrawl. We further filter the documents using the educational and STEM classifiers from ClassiCC-PT, which have been shown to improve continued pretraining of language models. The resulting corpus comprises roughly 12 billion tokens and is about six times larger than BrWaC, the corpus used to train BERTimbau.

For moBERTo, the model was trained for around 60 billion tokens, which corresponds to approximately 5 epochs over this dataset.

Languages

The dataset is monolingual Portuguese (pt), covering Brazilian and European variants as present in the FineWeb2 Portuguese subset.

Data Splits

The dataset is provided as a single pretraining split. No held-out validation or test split is defined, since evaluation is performed on external downstream benchmarks rather than on this corpus.

Dataset Creation

Source Data

The source is the Portuguese subset of FineWeb2, which is itself derived from CommonCrawl snapshots and processed with the FineWeb2 deduplication and quality pipeline.

Curation Procedure

Starting from the FineWeb2 Portuguese subset, we apply the educational and STEM classifiers from ClassiCC-PT to retain documents with higher educational and technical density. This filtering step targets content that benefits encoder pretraining for downstream tasks such as information retrieval, classification, named entity recognition, and natural language understanding. The final corpus contains approximately 12 billion tokens.

Considerations for Using the Data

Intended Use

This dataset is intended for pretraining and continued pretraining of Portuguese language models. It was used to train the moBERTo family of models.

Limitations and Biases

The corpus is derived from web text and reflects the biases, noise, and topical distribution present in CommonCrawl and in the educational and STEM filtering applied. The filtering favors educational and technical content, so the distribution is not representative of Portuguese language use in general.

Citation

@misc{laitz2026mobertomodernencoderportuguese,
      title={moBERTo: A Modern Encoder for Portuguese via Continued Pretraining of ModernBERT}, 
      author={Thiago Laitz and Thales Sales Almeida and João Guilherme Alves Santos and Giovana Kerche Bonás},
      year={2026},
      eprint={2606.22722},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2606.22722}, 
}