--- license: cc-by-nc-4.0 language: - en size_categories: - 1M *The SEC Filings Dataset: Reconstructing U.S. Corporate and Financial Disclosures into Layout-Faithful and Token-Efficient Pretraining Data.* Anonymous Authors (NeurIPS 2026 Evaluations & Datasets Track, under review). The full SFD corpus is estimated at ~500B tokens across ~18.4M filings (1994–present); this release is a public snapshot focused on the most recent four-and-a-half years. ## Key facts - **Format:** [MultiMarkdown](https://fletcher.github.io/MultiMarkdown-6/) — preserves merged-cell tables (`||` colspan, `^^` rowspan), indentation, and visual hierarchy that standard text extraction destroys. - **Source formats handled:** HTML (\~62%), XML (\~18%), plaintext (\~18%), SGML (\~2%), PDF-via-OCR (\~1%) — see paper §3. - **Filings:** ~3.4M parsed `.md` documents across 350+ filing types (10-K, 10-Q, 8-K, Form 4, N-PORT, 13F, 485BPOS, ABS-EE, …). - **License:** CC-BY-NC-4.0 (parsed). Underlying SEC filings are U.S. Government public domain. - **Storage:** Parquet shards with zstd-15 compression, one shard per (year, month). Token counts use the Qwen3-1.7B tokenizer. The headline ~152B figure counts the full released `parsed_md` field, including the prepended SEC/SGML filing metadata header. The regulatory body content alone is ~111B tokens; the difference is metadata such as CIK, accession, form type, filing date, period of report, and related SEC header fields. ## Schema Each row is one parsed filing. | field | type | description | |-------|------|-------------| | `accession` | string | SEC accession number (e.g. `0000320193-24-000123`) | | `file_stem` | string | Stem of the original submission file | | `year` | int16 | Filing year | | `month` | int8 | Filing month (1–12) | | `parsed_md` | string | MultiMarkdown reconstruction of the filing | | `char_count` | int64 | Character count of `parsed_md` | | `md5` | string | MD5 of `parsed_md` (UTF-8) | | `has_ocr` | bool | True if any portion required Mistral OCR | | `source_format` | string | Primary source format (`html`, `xml`, `plaintext`, `sgml`, `pdf`) | ## Loading ```python from datasets import load_dataset ds = load_dataset("anonymous-md/EDGAR_FILINGS_DATASET", split="train", streaming=True) for row in ds.take(1): print(row["accession"], row["year"], row["month"], len(row["parsed_md"])) ``` For a full local materialization (~50 GB on disk): ```python ds = load_dataset("anonymous-md/EDGAR_FILINGS_DATASET", split="train") ds = ds.filter(lambda r: r["year"] == 2024) # or by month, source_format, etc. ``` To recover the canonical SEC EDGAR URL for any row: ```python acc = row["accession"] # e.g. "0000320193-24-000123" acc_clean = acc.replace("-", "") url = f"https://www.sec.gov/Archives/edgar/data/{int(acc.split('-')[0])}/{acc_clean}/" ``` ## Sample subset (for reviewer inspection) Because the full corpus exceeds 4 GB, a small sample is provided at `sample/sfd-v1-sample.parquet` to allow quick inspection of data quality without materializing the full dataset. - **Path:** `sample/sfd-v1-sample.parquet` - **Sampling methodology:** stratified by year — first/last representative month of each year in the release (2022-01, 2023-01, 2024-01, 2025-01, 2025-06), with the first 25 rows from each, sorted by accession. This surfaces a mix of source formats (HTML / XML / SGML / PDF-OCR) and form types (Form 4, 10-Q, 8-K, NPORT-P, etc.) representative of the full corpus. - **Loading:** `pq.read_table("sample/sfd-v1-sample.parquet")` with the same schema as the full dataset. ## Methodology summary SFD treats filings as 2-D rendered grids rather than DOM trees: - **HTML:** Reconstructs the *visual* coordinate system; collapses the "Three-Column Hack" (\$ / value / closing-paren split into separate cells); coalesces fragmented multi-row headers using `border-*` and `margin-*` cues; preserves indentation hierarchy by binning CSS units into discrete ` ` levels. - **XML:** Routes 33 specialized schemas (Forms 3/4/5, 13F, N-PORT, N-CEN, etc.) through schema-aware emitters that reconstruct human-readable disclosures from field hierarchies. - **Plaintext / SGML:** Wraps fixed-width legacy filings in code fences to preserve column alignment; collapses 3+ blank lines to 2. - **PDFs:** Run through Mistral OCR 3 in 10-page batches; HTML table fragments converted to MMD; near-blank pages skipped via pixel-variance filter. Every row is prepended with ``-derived metadata (CIK, SIC, filing type, period of report, etc.) so each filing is self-contained. See paper §3 for full details. ## Companion benchmarks Two evaluation benchmarks are derived from SFD and reported alongside this dataset: - **EDGAR-OCR** — 300 hand-selected SEC tables, synthetically perturbed for contamination resistance, scored by adjusted recall over (content × placement × inline formatting). - **EDGAR-Forecast** — 50 companies × 5 numeric targets each (250 total) drawn from 2026 10-Q filings, evaluated agentically with prior 5-year filing history visible. Released separately when scoring is finalized. ## Citation ```bibtex @inproceedings{anonymous2026sfd, title={The SEC Filings Dataset: Reconstructing U.S. Corporate and Financial Disclosures into Layout-Faithful and Token-Efficient Pretraining Data}, author={Anonymous Authors}, booktitle={Advances in Neural Information Processing Systems (NeurIPS), Evaluations \& Datasets Track}, year={2026} } ``` ## License & terms - **Parsed corpus (this release):** [CC BY-NC 4.0](https://creativecommons.org/licenses/by-nc/4.0/) — non-commercial use only with attribution. Derivative works must keep this notice. - **Underlying raw filings:** U.S. Government public domain, available canonically from the SEC EDGAR system at . Redistribution of raw filings under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/) attribution-style metadata is permitted. - **Mistral OCR outputs** (the small PDF subset, `has_ocr == True`) are subject to the [Mistral AI Terms of Service](https://mistral.ai/terms/) at the time of generation; downstream redistribution within this CC BY-NC 4.0 corpus is permitted. ## Ethics, privacy, limitations - All filings are public regulatory disclosures with no expectation of privacy. - The dataset preserves filer-supplied content verbatim; SFD does **not** correct factual or accounting errors in the source filings. - The MMD reconstruction is high-fidelity but not perfect; estimated ~99% structural/semantic accuracy. A small minority of filings (notably highly visual exhibits with low OCR-recoverable content) may have degraded representation. - Token counts reflect the Qwen3-1.7B tokenizer; other tokenizers will differ. ## Provenance & versions - **v1** (this release): 2022-01 → 2025-06, parsed by SFD pipeline rev `sec_parser`. - Future releases will extend coverage to 1994–2021 and incrementally to 2025-07+. ## Contact - Anonymous Authors