Dataset

edgar-corpus

A reconstructed EDGAR corpus of company filings in English (Apache 2.0): a large sample of annual and quarterly reports and exhibits, hundreds of thousands of records. Comparable to other public SEC text sets and a solid starting point for filings-based research.

Open on Hugging Face

Data preview

Subset

Loading preview...

Dataset card

edgar-corpus is a text dataset of SEC annual report filings from 1993 to 2020, distributed by c3po-ai on the Hugging Face Hub. It is built from EDGAR-CRAWLER and is intended for financial natural language processing research.

What this dataset contains

The dataset aggregates English-language annual reports (10-K filings) of publicly listed companies sourced from the SEC EDGAR system, covering filing years from 1993 through 2020. It is provided as a raw, unannotated corpus with roughly 311,461 rows across train, validation, and test splits. Licensing is described as Apache-2.0 on the dataset card, while the underlying EDGAR data itself is publicly available.

Structure and fields

Each row is keyed by filename, cik, and year as string identifiers. The filing text is decomposed into standardized 10-K item columns including section_1, section_1A, section_1B, section_2, section_3, section_4, section_5, section_6, section_7, section_7A, section_8, section_9, section_9A, section_9B, section_10, section_11, section_12, section_13, section_14, and section_15. Users can load the full corpus or individual year configurations such as year_1993.

How analysts use it

Because the corpus exposes itemized risk, business, and MD&A text alongside cik and year identifiers, it is suited for building language models and text-mining pipelines over regulatory filings. Investment analysts can prototype retrieval, summarization, or sentiment and topic features tied to specific cik-year observations, while recognizing the authors flag that further cleaning is needed for downstream performance.

Documentation & references 1