{ "@context": { "@language": "en", "@vocab": "https://schema.org/", "citeAs": "cr:citeAs", "column": "cr:column", "conformsTo": "dct:conformsTo", "cr": "http://mlcommons.org/croissant/", "rai": "http://mlcommons.org/croissant/RAI/", "data": { "@id": "cr:data", "@type": "@json" }, "dataType": { "@id": "cr:dataType", "@type": "@vocab" }, "dct": "http://purl.org/dc/terms/", "examples": { "@id": "cr:examples", "@type": "@json" }, "extract": "cr:extract", "field": "cr:field", "fileProperty": "cr:fileProperty", "fileObject": "cr:fileObject", "fileSet": "cr:fileSet", "format": "cr:format", "includes": "cr:includes", "isLiveDataset": "cr:isLiveDataset", "jsonPath": "cr:jsonPath", "key": "cr:key", "md5": "cr:md5", "parentField": "cr:parentField", "path": "cr:path", "recordSet": "cr:recordSet", "references": "cr:references", "regex": "cr:regex", "repeated": "cr:repeated", "replace": "cr:replace", "sc": "https://schema.org/", "separator": "cr:separator", "source": "cr:source", "subField": "cr:subField", "transform": "cr:transform" }, "@type": "sc:Dataset", "name": "SFD-v1", "description": "The SEC Filings Dataset (SFD-v1) is an open, layout-faithful reconstruction of U.S. SEC EDGAR filings into token-efficient MultiMarkdown. This release covers ~3.8M filings filed between January 2022 and June 2025, totaling ~152B Qwen3-1.7B tokens, parsed from heterogeneous HTML, XML, plaintext, SGML, and PDF source formats. Tables are reconstructed using MultiMarkdown span notation (`||` for column spans, `^^` for row spans), preserving merged cells, multi-row headers, and indentation hierarchies that standard text extraction destroys. SFD-v1 supports long-context language modeling, financial reasoning, document understanding, and evaluation. The full SFD corpus is estimated at ~500B tokens across ~18.4M filings (1994-present); this release is a public snapshot of the most recent 4.5 years.", "conformsTo": [ "http://mlcommons.org/croissant/1.1", "http://mlcommons.org/croissant/RAI/1.0" ], "license": "https://creativecommons.org/licenses/by-nc/4.0/", "url": "https://huggingface.co/datasets/sfd-anonymous/sfd-v1", "version": "1.0.0", "datePublished": "2026-05-06", "creator": [ { "@type": "Person", "name": "Anonymous Authors", "affiliation": { "@type": "Organization", "name": "Anonymous Organization" } } ], "publisher": { "@type": "Organization", "name": "sfd-anonymous" }, "citeAs": "@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 and Datasets Track}, year={2026}}", "keywords": [ "finance", "sec", "edgar", "regulatory disclosure", "10-K", "10-Q", "8-K", "long-context", "table understanding", "multimarkdown", "pretraining", "financial reasoning" ], "isLiveDataset": false, "inLanguage": "en", "distribution": [ { "@type": "cr:FileObject", "@id": "repo", "name": "repo", "description": "Hugging Face dataset git repository hosting SFD-v1.", "contentUrl": "https://huggingface.co/datasets/sfd-anonymous/sfd-v1", "encodingFormat": "git+https", "sha256": "https://github.com/mlcommons/croissant/issues/80" }, { "@type": "cr:FileSet", "@id": "parquet-shards", "name": "parquet-shards", "description": "Per-month Parquet shards of parsed filings (zstd-15 compressed).", "containedIn": { "@id": "repo" }, "encodingFormat": "application/vnd.apache.parquet", "includes": "data/parsed/*.parquet" } ], "recordSet": [ { "@type": "cr:RecordSet", "@id": "filings", "name": "filings", "description": "One record per parsed SEC filing.", "field": [ { "@type": "cr:Field", "@id": "filings/accession", "name": "accession", "description": "SEC accession number (e.g. 0000320193-24-000123).", "dataType": "sc:Text", "source": { "fileSet": { "@id": "parquet-shards" }, "extract": { "column": "accession" } } }, { "@type": "cr:Field", "@id": "filings/file_stem", "name": "file_stem", "description": "Stem of the original submission file.", "dataType": "sc:Text", "source": { "fileSet": { "@id": "parquet-shards" }, "extract": { "column": "file_stem" } } }, { "@type": "cr:Field", "@id": "filings/year", "name": "year", "description": "Filing year (UTC).", "dataType": "sc:Integer", "source": { "fileSet": { "@id": "parquet-shards" }, "extract": { "column": "year" } } }, { "@type": "cr:Field", "@id": "filings/month", "name": "month", "description": "Filing month, 1-12.", "dataType": "sc:Integer", "source": { "fileSet": { "@id": "parquet-shards" }, "extract": { "column": "month" } } }, { "@type": "cr:Field", "@id": "filings/parsed_md", "name": "parsed_md", "description": "Layout-faithful MultiMarkdown reconstruction of the filing, with ``-derived metadata prepended.", "dataType": "sc:Text", "source": { "fileSet": { "@id": "parquet-shards" }, "extract": { "column": "parsed_md" } } }, { "@type": "cr:Field", "@id": "filings/char_count", "name": "char_count", "description": "UTF-8 character count of parsed_md.", "dataType": "sc:Integer", "source": { "fileSet": { "@id": "parquet-shards" }, "extract": { "column": "char_count" } } }, { "@type": "cr:Field", "@id": "filings/md5", "name": "md5", "description": "MD5 hash of UTF-8 encoded parsed_md.", "dataType": "sc:Text", "source": { "fileSet": { "@id": "parquet-shards" }, "extract": { "column": "md5" } } }, { "@type": "cr:Field", "@id": "filings/has_ocr", "name": "has_ocr", "description": "True if any portion of the filing required Mistral OCR (PDF or image content).", "dataType": "sc:Boolean", "source": { "fileSet": { "@id": "parquet-shards" }, "extract": { "column": "has_ocr" } } }, { "@type": "cr:Field", "@id": "filings/source_format", "name": "source_format", "description": "Primary source format detected for the filing: html, xml, plaintext, sgml, or pdf.", "dataType": "sc:Text", "source": { "fileSet": { "@id": "parquet-shards" }, "extract": { "column": "source_format" } } } ] } ], "rai:dataCollection": "Filings were retrieved from the public U.S. SEC EDGAR system (https://www.sec.gov/edgar) via canonical archive URLs. Submission types span 350+ form variants. Each submission's primary disclosure document was parsed; auxiliary attachments and exhibits were partitioned via SGML tags and parsed independently when content-bearing.", "rai:dataPreprocessingProtocol": "Format-routed parsing: HTML via 2-D coordinate-system reconstruction; XML via 33 schema-aware emitters; plaintext/SGML preserved via code-fence wrapping; PDFs via Mistral OCR 3 (10-page batches, near-blank-page filtering). Numerical normalization removes commas; non-semantic artifacts (orphan page numbers) are dropped; SEC header metadata (CIK, SIC, form type, period of report) is prepended to each document.", "rai:dataAnnotationProtocol": "No human annotation. Source-format labels and OCR-usage flags are assigned automatically by the parser.", "rai:dataReleaseMaintenancePlan": "Versioned releases via the HuggingFace dataset repository. v1 covers 2022-01 to 2025-06. Future versions will (a) backfill 1994-2021, (b) extend coverage past 2025-06 monthly. Issues should be filed at the HF dataset repo discussions.", "rai:dataUseCases": "Long-context language model pretraining on financial regulatory disclosures; financial reasoning, forecasting, and compliance benchmarks; document layout understanding and table parsing; retrieval-augmented generation over financial filings; XBRL-aligned named-entity and relation extraction.", "rai:dataLimitations": "Parsing fidelity is high (~99% estimated structural/semantic accuracy) but not perfect. A small minority of highly visual exhibits with poor OCR recovery may have degraded representation. Token counts reflect the Qwen3-1.7B tokenizer and will differ for other tokenizers. Coverage is limited to 2022-01 to 2025-06 in this release. Users should not treat SFD-derived outputs as financial, legal, compliance, or investment advice, and should account for parser/OCR errors before using the data in decision workflows.", "rai:personalSensitiveInformation": "All filings are public regulatory disclosures filed by issuers and reporting persons under SEC rules; no expectation of privacy applies to their content. Reporting-person personal addresses and identifiers (e.g., insider ownership filings) are present as part of the regulatory record.", "rai:annotationsPerItem": "0 (no human annotation)", "rai:annotatorDemographics": "N/A - automated parsing only.", "rai:dataSocialImpact": "Lowers the barrier to studying U.S. corporate and financial disclosure at scale, supports auditing of LLMs on real-world regulatory text, and provides an open alternative to closed financial corpora.", "rai:dataBiases": "Coverage is U.S.-centric and skewed toward issuers active in U.S. capital markets. Filing-type token distribution is highly non-uniform (the top 5 form types account for ~35% of tokens). Recent years are over-represented relative to the full EDGAR archive in this v1 release.", "alternateName": "sfd-v1", "sameAs": "https://huggingface.co/datasets/sfd-anonymous/sfd-v1" }