Dataset

SEC-EDGAR

Text of SEC filings pulled from EDGAR spanning millions of records in English (Apache 2.0): 10-K, 10-Q, 8-K, annual reports and exhibits, tagged for text-generation and classification. A practical base corpus for earnings- and filings-driven signals, from risk factors and MD&A to management commentary.

Open on Hugging Face

Data preview

Subset

Loading preview...

Dataset card

SEC-EDGAR is a large-scale text dataset of corporate filings scraped from the U.S. SEC's EDGAR system by TeraflopAI, Datamule, and Eventual, totaling roughly 590 GB across 8 million samples. It provides raw and parsed plaintext of filings together with structured filing and filer metadata for machine-learning research.

What this dataset contains

The release covers major filings drawn from the SEC EDGAR database, with 8,055,455 rows in the published train split and a stated size band of 1M–10M samples. Each row stores raw filing contents alongside HTML/XML plaintext that was extracted with selectolax and the doc2dict and secsgml libraries. Document and filer metadata accompany every sample, allowing linkage back to EDGAR submission records.

Structure and fields

Rows are organized as string-valued columns named content, text, metadata_accession-number, metadata_filing-date, metadata_documents, metadata_period, and metadata_filer. The content field carries the raw submission bytes, while text holds the parsed plaintext output. The metadata_accession-number, metadata_filing-date, metadata_period, and metadata_filer fields provide identifiers, dates, the reporting period, and the filing entity (including Central Index Key and company name), and metadata_documents captures per-document attributes such as file type, sequence, filename, description, and SGML byte counts.

How analysts use it

The combination of plaintext and rich filing metadata makes it a base layer for text-generation and text-classification tasks tagged by the publisher, such as training language models or classifiers on 10-K and similar disclosures. Investment teams can use the accession, filing-date, and CIK fields to join samples to internal universes for event-driven backtests, factor construction, or NLP-based sentiment and risk-signal research on public-company fundamentals. License terms beyond the Hugging Face Apache-2.0 tag are not specified in the supplied facts, so commercial reuse should be reviewed against SEC redistribution rules.

Documentation & references 1