Dataset

SEC-EDGAR

A joint release by Datamule, Teraflop AI, and Eventual provides the SEC-EDGAR dataset, comprising 590 GB of content drawn from all major filings in the SEC EDGAR database and totaling roughly 8 million samples and 43 billion tokens. The data was assembled using the datamule-python library and the official datamule API developed by John Friedman, with datamule serving as a Python package for collecting and manipulating such records.

Open on Hugging Face

Data preview

Subset

Loading preview...

Dataset card

SEC-EDGAR is a large-scale text dataset of filings scraped from the SEC EDGAR system, distributed via a community repository. It packages raw and parsed plaintext of major filings alongside accession-level metadata for roughly eight million records.

What this dataset contains

The release spans approximately 8 million individual filings and 590 GB of content drawn from all major SEC EDGAR submission types. Each record exposes both the raw document bytes (HTML, XML, SGML, and embedded PDFs) and a plaintext extraction produced with selectolax, doc2dict, and secsgml. Repository tags describe the size as 1M–10M rows in English, oriented toward text-generation and text-classification tasks.

Structure and fields

The schema exposes a `content` string with the raw filing payload and a `text` string with parsed plaintext. Filing-level metadata is broken into `metadata_accession-number`, `metadata_filing-date`, `metadata_period`, and `metadata_documents` (a JSON-encoded list covering document type, sequence, description, filename, and SGML byte count), plus `metadata_filer` carrying company name and Central Index Key information. The single provided split is a `train` partition of about 8,055,455 rows.

How analysts use it

Quantitative and NLP teams can use the corpus to build language models or classifiers over regulatory disclosures, joining `metadata_filer` and `metadata_filing-date` to issuer histories. Long/short and event-driven funds could train document-level models on 10-K, 10-Q, and 8-K texts to surface risk language or sentiment around specific periods. License and publisher status are unstated in the listing, so commercial use should be confirmed against the underlying source terms before deployment.

Documentation & references 1