Dataset

SEC-EDGAR

A large-scale release produced jointly by Datamule, Teraflop AI, and Eventual covering roughly 590 GB of SEC EDGAR content. It comprises about eight million filings totaling 43 billion tokens, gathered using the datamule-python package and the official datamule API.

Open on Hugging Face

Data preview

Subset

Loading preview...

Dataset card

SEC-EDGAR is a large-scale text corpus of corporate filings released by baridhi in collaboration with Datamule, Teraflop AI, and Eventual. It covers roughly 8 million major SEC submissions with parsed plaintext and filing metadata, intended for training language models on financial disclosure text.

What this dataset contains

The corpus mirrors filings from the SEC EDGAR system, spanning all major submission types and totaling about 590 GB and 43 billion tokens across roughly 8 million samples. Each record combines the raw document payload with extracted plaintext and structured metadata. Underlying data is collected using the datamule-python library and the official datamule API, subject to the EDGAR rate limit of 10 requests per second.

Structure and fields

The dataset is delivered as a single training split of 8,055,455 rows, tagged for text generation and text classification tasks in English with an Apache-2.0 license. Each row carries the raw content in content, a plain-text rendering in text, and several metadata string fields: metadata_accession-number, metadata_filing-date, metadata_period, metadata_documents (covering file type, sequence, filename, description, and SGML byte count), and metadata_filer (covering company name and Central Index Key). Parsing of the SGML, HTML, and XML submissions is performed with selectolax alongside modified doc2dict and secsgml libraries.

How analysts use it

Quantitative researchers can use the parsed text and metadata to build filings-based datasets for language modeling, classification of disclosure types, or extraction of structured items from 10-K and 10-Q narratives. The filer and period fields support joining filings to company identifiers and reporting dates for event studies and back-tests. Because the release is a researcher-curated mirror of public EDGAR data, it is best suited as a training corpus or offline archive rather than a low-latency feed.

Documentation & references 1