Dataset

SEC-EDGAR

Datamule, Teraflop AI, and Eventual jointly contributed a 590 gigabyte SEC EDGAR collection encompassing about 8 million filings and 43 billion tokens, assembled using the datamule-python library 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 roughly 8 million filings from the U.S. Securities and Exchange Commission's EDGAR system, distributed on the Hugging Face hub by publisher adityaag2k. It packages raw HTML/XML, parsed plaintext, and structured filing metadata in a single training-ready table of about 8 million rows.

What this dataset contains

Each row corresponds to a single SEC filing drawn from the major submissions on EDGAR, with the original document body preserved alongside a clean plaintext rendering. The collection was assembled using the datamule-python library and the datamule API, and plaintext was extracted with the selectolax parser plus modified versions of doc2dict and secsgml. Per the project description, the release totals roughly 590 GB of data and around 43 billion tokens.

Structure and fields

The published schema is a single tabular split (train, 8,055,455 rows) with seven string columns: content holds the raw HTML/XML payload, text holds the parsed plaintext, and metadata_accession-number, metadata_filing-date, metadata_documents, metadata_period, and metadata_filer carry filing-level identifiers and dates. The metadata_documents field embeds per-document attributes such as form type, sequence, description, and SGML byte count, while metadata_filer embeds the filer block including company name and Central Index Key.

How analysts use it

The corpus is shaped for text-generation and text-classification workloads over regulatory filings, making it suitable for training or fine-tuning models that summarize 10-Ks and 10-Qs, classify filing types, or extract financial disclosures. Because every record retains the accession number, filing date, and reporting period, outputs can be joined back to issuer-level time series for event studies, earnings-sentiment tracking, or automated earnings-preview generation within an investment workflow.

Documentation & references 1