stanford-edgar-parser 0.1.1

Layout-faithful SEC EDGAR filing parser from the Stanford EDGAR Filings Dataset.

Stanford EDGAR Filings Dataset Parser

The Stanford EDGAR Filings Dataset (SEFD) is a 550B-token reconstruction of 18.5M SEC filings into layout-faithful MultiMarkdown for LLM pretraining, financial reasoning, document understanding, and evaluation.

SEFD reverse-engineers EDGAR's heterogeneous disclosure formats into a token-efficient representation that preserves financial tables, indentation, merged headers, numeric signs, currency and percent symbols, document hierarchy, and other layout cues that carry financial meaning. Internal validation shows our rule-based reconstruction methodology achieves greater than 99% structural and semantic accuracy on sampled outputs.

The software routes and parses the full EDGAR source-format surface, including legacy fixed-width text, tag-soup HTML, SGML wrappers, XML submissions, and PDF attachments, with specialized reconstruction for more than 30 SEC XML schemas, including Forms 3, 4, 5, D, 13D/G, N-PX, N-PORT, N-CEN, 13F, 144, ATS-N, 1-A/K/Z, C, MA, TA, X-17A-5, 24F-2NT, ABS-EE, and related amendments, withdrawals, and corrections. PDF attachments are parsed with Mistral OCR 3.

Our first public release, SEFD-v1, is a 152B-token dataset covering filings from January 2022 through June 2025.

• Paper: The Stanford EDGAR Filings Dataset
• Dataset: SEFD-v1 on Hugging Face

Install

From PyPI:

pip install stanford-edgar-parser

Or directly from GitHub:

pip install "stanford-edgar-parser @ git+https://github.com/Stanford-Advanced-FinTech-Lab-SAFTL/stanford-edgar-filings-dataset.git"

Usage

Parse a local filing and convert tables to MultiMarkdown:

stanford-edgar-parser path/to/filing.txt --to_mmd

or:

python -m stanford_edgar_parser path/to/filing.txt --to_mmd

Optional rendering helpers:

node multimarkdown.js path/to/file.md > file.html
node html-to-pdf.mjs file.html file.pdf

Agent Skills

Install bundled Codex and Claude skills:

stanford-edgar-install-skill

or:

from stanford_edgar_parser.ai import install_skill

install_skill()

MCP

Package install:

[mcp_servers.stanford_edgar_parser]
command = "uvx"
args = ["--from", "stanford-edgar-parser", "stanford-edgar-mcp"]
startup_timeout_sec = 120

GitHub install:

[mcp_servers.stanford_edgar_parser]
command = "uvx"
args = [
  "--from",
  "stanford-edgar-parser @ git+https://github.com/Stanford-Advanced-FinTech-Lab-SAFTL/stanford-edgar-filings-dataset.git",
  "stanford-edgar-mcp"
]
startup_timeout_sec = 120

The package-installed MCP server always exposes parse_filing. Repo-local rendering and review tools are exposed when the full clone includes multimarkdown.js, html-to-pdf.mjs, and tools/.

Citation

@article{bettencourt2026stanfordedgar,
  title={The Stanford EDGAR Filings Dataset: Reconstructing U.S. Corporate and Financial Disclosures into Layout-Faithful and Token-Efficient Pretraining Data},
  author={Bettencourt, Nick and Ding, Xiaowei and Giesecke, Kay},
  journal={arXiv preprint arXiv:2606.18192},
  year={2026},
  url={https://arxiv.org/abs/2606.18192}
}