Dataset

edgar-corpus

A corpus of SEC EDGAR filings in English (Apache 2.0), roughly a few hundred thousand company documents including annual and quarterly reports and exhibits. The filings text is organized for model pretraining and retrieval, but the same documents are the raw material for filings-based and disclosure signals.

Open on Hugging Face

Data preview

Subset

Loading preview...

Dataset card

EDGAR-CORPUS is a research dataset of annual reports filed by public companies with the SEC EDGAR system, covering 1993 through 2020 and released by the authors of a 2021 ECONLP workshop paper. It is distributed as raw, unannotated English text intended for financial natural language processing work.

What this dataset contains

The dataset bundles the textual content of SEC EDGAR annual report filings submitted by public firms between 1993 and 2020, produced by the same group that authored the EDGAR-CORPUS paper. It is supplied in two configurations: a single 'full' configuration holding the entire corpus, and per-year configurations such as 'year_1993' that allow analysts to pull a specific filing year. The Hugging Face listing indicates a size band of 100K to 1M examples, with English-only content and no annotation layer.

Structure and fields

Filings can be loaded either as a whole via the 'full' configuration or per calendar year via configurations named like 'year_1993', each exposed with a 'train' split. The published schema does not enumerate explicit column names; per-row content corresponds to the underlying EDGAR filing text rather than a fixed tabular layout. No sample rows are exposed in the dataset card, so specific field labels beyond the configuration and split identifiers cannot be confirmed.

How analysts use it

Because the corpus is raw filing text with no labels, it is suited to pretraining and benchmarking financial language models, topic modeling on 10-K disclosures, and dictionary-based extraction of risk or governance language across long windows. Investment teams could in principle mine the 1993-2020 filings to track disclosure trends or build features for event studies, though any such use would require downstream labeling work given the absence of target variables. Licensing terms beyond the underlying SEC filings and the Apache-2.0 tag noted in the metadata are not clearly stated on the dataset card.

Documentation & references 1