The Cocoruta Hub: Open and Curated Corpora, Datasets and Language Models on Brazilian Ocean Law

This paper describes Cocoruta, an open-source hub for computational linguistics resources and language models related to the Brazilian legal context, with a focus on ocean law. The Cocoruta hub consists of open-access, curated corpora and datasets, and fine-tuned language models. Cocoruta includes two sets of resources. The first set is associated with a training dataset for question-answering tasks, whereas the second set features a dataset prepared to support model training for dialogue tasks, in addition to having refined curation procedures. In this paper, we provide a comprehensive analysis of Cocoruta's contributions, including information to ensure transparency and reproducibility of Cocoruta's construction process, fine-tuning of language models, and quantitative and qualitative evaluations of models. Quantitative evaluations are based on various performance metrics, while qualitative evaluations are conducted using human assessment procedures. This work contributes by advancing fundamental resources in specialized language domains and by fostering research and development in Brazilian legal natural language processing, serving as a hub to bring together efforts in this field.