# The Material Contracts Corpus: Codebook

The Material Contracts Corpus is a dataset of exhibits filed as “material contracts” with the U.S. Securities and Exchange Commission (SEC), obtained through the publicly available SEC EDGAR database. This codebook describes the metadata that accompanies the contracts.

### Columns

Each metadata row describes one contract exhibit. The metadata is available as CSV (with a header row) or Apache Parquet; both formats carry the identical columns below, in this order.

| Column | Type | Description |
| --- | --- | --- |
| `year` | integer | Filing year. |
| `quarter` | string | Filing quarter, "QTR1" through "QTR4". |
| `cik` | string | Filer Central Index Key (CIK) as it appears in EDGAR. |
| `accession` | string | SEC accession number, digits only, as it appears in EDGAR. |
| `company_name` | string | Filer name as it appears in EDGAR; empty if unmatched. |
| `form_type` | string | Submission form type (e.g. "10-K" or "8-K"); empty if unmatched. |
| `date_filed` | date | Filing date (YYYY-MM-DD format); empty if unmatched. |
| `contract_link` | string | Stable URL of the contract on this site. |
| `exhibit` | string | Exhibit type parsed from the filename (e.g. "EX-10.1"); empty if unparseable. |
| `exhibit_lead` | string | The lead part of `exhibit` (e.g. "EX-10"); empty if unparseable. |
| `content_type` | string | MIME type of the original document, derived from its filename extension. |
| `label` | string | Machine-generated contract category (see "Contract labels" below). |
| `parties` | list of strings | Machine-identified contracting parties, resolved to entity master names (see "Contracting parties" below). |
| `parties_raw` | list of strings | The extracted party names as they appear in the document before entity resolution, plus the filer name from EDGAR. |
| `image_count` | integer | Number of `<img>` tags in the original document markup. |
| `ocr_labeled` | boolean | True if OCR-derived text from the document's images was in view when the label was assigned. |
| `amend` | boolean | Document-type flag for amendments. |
| `restate` | boolean | Document-type flag for restatements. |
| `joinder` | boolean | Document-type flag for joinders. |
| `termination` | boolean | Document-type flag for terminations. |
| `archive_name` | string | The contract's path inside a downloaded archive, under `archive/documents/`. |

### Contract labels

The `label` column is a machine-generated contract category, one of: `purchase`, `employment`, `security`, `other`, `services`, `na`, `lease`, `shareholder`. Labels are assigned by a text classifier that reads the first 6,000 characters of each document's extracted text, so they may contain inaccuracies. The label `na` means the document was not judged to be a substantive contract. The taxonomy follows Adelson and Nyarko (2025), with differences noted under “Relationship to Paper” below.

### Document-type flags

The columns `amend`, `restate`, `joinder`, `termination` are non-exclusive boolean flags. A flag is true if and only if its keyword stem appears in the first 30 words of the document's extracted text — deterministic string matching, no model involved.

### Contracting parties

In addition to contract labels and document-type flags, this corpus enumerates the parties involved in each contract. The process to identify parties is twofold: first, a Named Entity Recognition (NER) model is used to extract party names from the document's opening text. Then, a deterministic fuzzy matching algorithm is used to de-duplicate and link matching parties (e.g., "jpmorgan chase" and "jpmorgan chase & co" would likely be linked together). This algorithm also determines a *master name* for each party, chosen primarily by the number of documents in which that name appears.

A contract's metadata contains two columns relating to its parties: `parties` and `parties_raw`. `parties_raw` is the list of unique party names extracted from the document's opening text by the NER model plus the filer name from EDGAR, which may contain semantic duplicates. `parties` is the list of master names identified by the fuzzy matching algorithm.

Note: Party identification is both predictive and heuristic-based, so it may contain errors. Additionally, since they are computed primarily by per-document frequencies, *master names* may change between releases as the corpus grows.

### Archive layout

A full download unpacks to a single `archive/` folder holding the metadata file and every contract at `archive/documents/<cik>/<accession>/<filename>`, such that a row's `archive_name` column is exactly that path under `archive/documents/`. Downloads offer three treatments of the images some contracts reference: `transcribed` replaces each `<img>` tag inline with the image's OCR text, `original` ships the referenced image files beside their document so relative links resolve, and `none` ships the documents alone.

### Relationship to Paper

This corpus continues the dataset described in the paper, with the following differences.

- The paper analyzes filings from 2000 through 2023, while the corpus has no fixed end date and automatically updates each quarter.
- The heuristics that filter material contracts from EDGAR filings have been improved.
- The second phase of party identification (fuzzy matching to link and de-duplicate) has been improved.
- The paper did not use Optical Character Recognition (OCR) to extract text from images, so contracts with image-dominant content were generally labeled `na`. Roughly a tenth of contracts reference images, and a subset of those now carry a different label; rows with `ocr_labeled` true were labeled with OCR text in view, so excluding them approximates the paper's conditions.
- The `purchase/m&a` category has been renamed to `purchase` in the metadata. It is still shown as "Purchase / M&A" on the website.

### Citation & license

Adelson, Peter, and Nyarko, Julian. *The Material Contract Corpus.* arXiv preprint arXiv:2504.02864, 2025. https://arxiv.org/abs/2504.02864

The corpus is licensed under CC BY-NC-SA 4.0, which permits non-commercial use with attribution. For commercial use, contact jnyarko@law.stanford.edu. The corpus lives at https://mcc.law.stanford.edu.