datasheet.md

# Datasheet: SFD — SEC Filings Dataset (v1)

Following the *Datasheets for Datasets* template (Gebru et al., 2021).

## Motivation

**For what purpose was the dataset created?**
SFD-v1 was created to provide a large, open, layout-faithful corpus of U.S. SEC EDGAR regulatory filings for long-context language modeling, financial reasoning research, document understanding, and contamination-resistant evaluation. Existing financial corpora (BloombergGPT/FinPile, BeanCounter, EDGAR-CORPUS) are either proprietary, narrowly scoped, or strip the visual table structure that financial reasoning depends on.

**Who created the dataset?**
Anonymous Authors (Anonymous Organization).

**Who funded the creation?**
Anonymous Organization.

## Composition

**What do the instances represent?**
Each instance is one parsed SEC EDGAR filing, represented as MultiMarkdown text plus metadata fields (accession number, year/month, source format, OCR flag, char count, MD5). Filings span 350+ form types including 10-K, 10-Q, 8-K, Form 4, N-PORT, 13F, 485BPOS, ABS-EE, and many others.

**How many instances?**
Approximately 3.4M filings filed January 2022–June 2025. Final count provided in `metrics.json` upon dataset completion.

**Does the dataset contain all possible instances or a sample?**
This is a complete pull of EDGAR filings within the date range, with the following exceptions:
- Filings that fail parsing after 3+ retry attempts across configurations are excluded (logged as "challenging") — a small minority (<0.1%).
- Auxiliary exhibit attachments (e.g., individual XML information tables) are deduplicated against their parent submissions.

**What data does each instance consist of?**
Plain text in MultiMarkdown format (`parsed_md`), plus structured metadata fields. No raw bytes, images, or paper PDFs are included.

**Is there a label or target?**
No. The dataset is unlabeled regulatory text. Companion benchmarks EDGAR-OCR and EDGAR-Forecast (released separately) provide labeled evaluation tasks derived from this corpus.

**Is any information missing?**
- Filings before 2022-01 are not in this release.
- Filings after 2025-06 will be added in future releases.
- Approximately <0.1% of filings within the date range that failed all parsing retries are absent.

**Are relationships between instances explicit?**
Same-issuer relationships can be recovered via the SEC accession number's CIK (Central Index Key) prefix and the SEC-HEADER metadata embedded at the top of each `parsed_md`. The SEC EDGAR system is the canonical source for these relationships.

**Are there recommended data splits?**
No. SFD-v1 is intended as a single training corpus. Researchers wanting train/eval splits should:
- For temporal evaluation: hold out the latest month (2025-06).
- For company-level evaluation: hold out by CIK.
- For benchmark evaluation: use EDGAR-OCR and EDGAR-Forecast.

**Are there errors, sources of noise, or redundancies?**
- **Parser fidelity:** ~99% estimated structural/semantic accuracy. Residual errors include occasional misclassification of fragmented headers, misaligned multi-row tables in pathological HTML, and rare cases where the "Three-Column Hack" merger fails when adjacent columns contain footnote markers that resemble currency modifiers.
- **OCR fidelity:** Mistral OCR 3 was used for PDF content; small fraction (~1% of corpus) carries `has_ocr=True`. EDGAR-OCR benchmark measures fidelity at ~75%.
- **Filer-level errors:** SFD does NOT correct factual or accounting errors in the source filings.
- **Duplicates:** Some filings re-submit prior content as exhibits or amendments; deduplication by `md5` is left to the consumer.

**Is the dataset self-contained?**
Yes. Each row's `parsed_md` is self-contained and includes prepended SEC-HEADER metadata.

**Does the dataset contain confidential data?**
No. All filings are public regulatory disclosures.

**Does the dataset contain offensive, insulting, threatening, or anxiety-inducing content?**
No, the corpus is corporate regulatory disclosures. Trace amounts of legal proceeding text or risk-factor language could in theory describe sensitive events but are factual and public.

**Does the dataset relate to people?**
Some filings (e.g., Form 4, Form 13D/G) name natural persons in their capacity as insiders, beneficial owners, or filers. These individuals are public regulatory reporters; their personal information appears only in the form required by SEC rules.

**Does the dataset identify any subpopulations or contain protected categories?**
No.

## Collection Process

**How was the data acquired?**
By downloading filings via the SEC EDGAR public archive (https://www.sec.gov/edgar). All access conformed to SEC rate-limiting and User-Agent guidance.

**What mechanisms or procedures were used to collect the data?**
Custom Python downloader respecting SEC's 10-requests-per-second guidance, with periodic master-index reconciliation.

**If the dataset is a sample, what was the sampling strategy?**
Not a sample — exhaustive over the date range.

**Who was involved in the data collection?**
The authors and engineering team within the Anonymous Organization.

**Over what timeframe was the data collected?**
Filings cover 2022-01 through 2025-06. Collection (download + parse) was performed in 2024–2026 on shared HPC infrastructure.

**Were any ethical review processes conducted?**
The data is publicly filed regulatory disclosures, so no IRB review was required. The authors confirmed compliance with SEC's terms of service and data-access guidance.

## Preprocessing / Cleaning / Labeling

**Was any preprocessing done?**
Yes. The full SFD parsing pipeline:
1. SGML-header parsing → routing by `CONFORMED SUBMISSION TYPE`.
2. Format-specific reconstruction (HTML / XML / plaintext / SGML / PDF).
3. Numerical normalization (decomma).
4. Non-semantic artifact removal (orphan page numbers, etc.).
5. Metadata prepending (CIK, SIC, form type, period of report).

**Was the raw data saved alongside?**
Raw filings are not redistributed in this release but are available canonically from SEC EDGAR. Rows include `accession`, `file_stem`, and `md5` to support reproducibility.

**Is the preprocessing software available?**
The SFD parser source will be released alongside paper acceptance. See paper §3 for full methodological detail.

## Uses

**Has the dataset been used for any tasks already?**
- Long-context language model pretraining experiments by the dataset authors.
- Construction of EDGAR-OCR and EDGAR-Forecast benchmarks.

**What other tasks could the dataset be used for?**
- Financial named-entity / relation extraction (alignable with XBRL).
- Retrieval-augmented generation over filings.
- Compliance and audit automation.
- Long-context modeling research (10K-1M+ tokens per document).
- Table understanding research (tables faithfully preserved in MMD).

**Is there anything about the composition of the dataset that might impact future uses?**
- Recent-year skew may bias models toward post-pandemic disclosure language.
- The U.S.-centric scope limits cross-jurisdictional finance research.
- The CC BY-NC license restricts commercial use of the parsed corpus.

**Are there tasks for which the dataset should not be used?**
- Investment decisions or trading: SFD is a research dataset; do not use as a financial advisor.
- Identifying or harming individual reporting persons: while names are present, using the corpus to target individuals is contrary to its research intent.
- Commercial product training: forbidden by CC BY-NC 4.0.

## Distribution

**Will the dataset be distributed to third parties?**
Yes — publicly via HuggingFace Datasets at <https://huggingface.co/datasets/mikedd/EDGAR_FILINGS_DATASET>.

**How will the dataset be distributed?**
Parquet shards on HuggingFace, accessible via `datasets.load_dataset(...)`.

**When will the dataset be distributed?**
v1: with the NeurIPS 2026 E&D Track submission. Subject to date of camera-ready release.

**Will the dataset be distributed under a copyright or other IP license?**
Yes — CC BY-NC 4.0. See LICENSE notice in README.

**Have any third parties imposed IP-based or other restrictions?**
- Mistral AI Terms of Service govern OCR-derived content within the corpus (~1% of rows have `has_ocr=True`). Mistral permits downstream redistribution of OCR outputs.
- Filer-supplied content is U.S. Government public domain (filings are public records under SEC rules).

**Do any export controls or other regulatory restrictions apply?**
No.

## Maintenance

**Who is supporting/maintaining the dataset?**
The Anonymous Organization.

**How can the owner/curator be contacted?**
Anonymous during double-blind review. HuggingFace repo discussions for issue reports.

**Is there an erratum?**
None at v1 release; future errata will be appended to this datasheet and tagged in the HF repo.

**Will the dataset be updated?**
Yes — see versioning plan in README. Coverage will be extended both backward (1994–2021) and forward (2025-07+).

**If the dataset relates to people, are there limits on retention of data relating to specific individuals?**
N/A — all data is publicly filed regulatory information with no expectation of privacy.

**Will older versions be supported/hosted?**
Yes — version tags will remain accessible via HF dataset revisions.

**Mechanisms for contribution?**
External contributions of additional filing-type parsers, corrections to specific filings, or extension to non-U.S. regulatory archives can be proposed via HF discussions or by contacting the authors.