Skip to content

Dataset · heitorrosa

Brazilian regulator CVM financial report corpus

heitorrosa/cvm-corpus

Cleaned and chunked documents from Brazilian regulator CVM in Portuguese for language model fine-tuning.

What's inside

A corpus of regulatory documents from the Brazilian CVM commission in Portuguese (PT-BR), with a volume of 100K–1M records (7.2 GB in jsonl format). Texts are cleaned and split into fragments with metadata: company, CNPJ, category, topic, date, and extraction quality score. Besides the full archive, it includes a DSIR-filtered subset of 250M tokens for DAPT.

What you can do with it

  • Domain-adaptive pre-training (DAPT) of language models on financial Portuguese
  • Building RAG systems using Brazilian corporate and regulatory reports
  • Training models to extract metadata from financial documents
Tasks
Text generation, Fill mask
Languages
Portuguese
Size
100K–1M rows
Downloads this month
20K
Likes
0

Models trained on it

How to load it

from datasets import load_dataset

ds = load_dataset("heitorrosa/cvm-corpus")