Datasetpublic

heitorrosa/cvm-corpus

CVM Filings Corpus (PT-BR) Brazilian CVM regulatory filings in Portuguese, cleaned and chunked for language-model pretraining. Built for DAPT on financial Portuguese. Contents Path What output/corpus.jsonl Full corpus (7.2 GB). Chunk schema: text, company, cnpj, category, subject, date, year, document_id, chunk_id, extraction_quality. output/corpus-250M.jsonl DSIR-selected 250M-token subset (token count by chars÷4 proxy ≈ 150M whitespace tokens): 187… See the full description on the dataset page: https://huggingface.co/datasets/heitorrosa/cvm-corpus.

sourceHugging Facegpl-3.0updated 8d agoView on Hugging Face

0likes9.6kdownloads

Dataset Card

CVM Filings Corpus (PT-BR)

Brazilian CVM regulatory filings in Portuguese, cleaned and chunked for language-model pretraining. Built for DAPT on financial Portuguese.

Path	What
`output/corpus.jsonl`	Full corpus (7.2 GB). Chunk schema: `text, company, cnpj, category, subject, date, year, document_id, chunk_id, extraction_quality`.
`output/corpus-250M.jsonl`	DSIR-selected 250M-token subset (token count by chars÷4 proxy ≈ 150M whitespace tokens): 187,989 chunks, 1,030 companies, 2003–2020 unbroken.
`output/split-250M.json`	Selection manifest: per-category counts, token stats, KL pre/post.
`<year>/pdfs/`	Archival mirror of source PDFs. Years over 10k files are sharded as `part-*`; the jsonl above is the canonical artifact.

Pipeline

Scrape (CVM ENET) -> merge -> clean -> DSIR-select against an INPE-style target distribution.

Exact code at: https://github.com/mansa-team/musa

Data Flow

Data quality

Measured on the 250M subset: residual junk 3.7%, exact-duplicate 0.4%, median chunk 943 whitespace tokens.

TODO

—[ ] Add 2021–2026 filings and a final merge

Cite

@misc{heitor_de_oliveira_rosa_2026,
	author       = { Heitor de Oliveira Rosa },
	title        = { cvm-corpus (Revision bf07404) },
	year         = 2026,
	url          = { https://huggingface.co/datasets/heitorrosa/cvm-corpus },
	doi          = { 10.57967/hf/10644 },
	publisher    = { Hugging Face }
}