Research map / Large language models / Language, safety and interpretability

The SEC Filings Dataset: Reconstructing U.S. Corporate and Financial Disclosures into Layout-Faithful and Token-Efficient Pretraining Data

NeurIPS 2026 · Large language models › Language, safety and interpretability · approach: Data and evaluation methodology

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces SEFD, a layout-faithful MultiMarkdown reconstruction of the SEC EDGAR archive (152B-token public snapshot), plus EDGAR-OCR and EDGAR-Forecast benchmarks derived from it.

Problem

Clean long-context pretraining data is scarce, and EDGAR filings are noisy HTML/XML/PDF/plaintext whose table and layout structure is lost in standard extraction pipelines.

Approach

A visual-first, format-specific parsing pipeline that reconstructs rendered layout (indentation, split numeric cells, merged headers) and serializes filings into token-efficient MultiMarkdown, with spans encoded by pipes and carets.

Key results

• GPT-5.5 achieves 51.8% on EDGAR-Forecast (250 targets, 50 companies), the highest among tested models.
• On EDGAR-OCR, Qwen3.6-35B-A3B scores 75.78% adjusted recall, with Mistral OCR 3 at 75.33% at about one-eleventh the median latency.
• SEFD MMD reaches 94.5% adjusted recall on HTML table reconstruction versus 75.7% for EdgarTools and 70.4% for to_markdown.

Surprising findings

• Qwen3.6-35B-A3B slightly outperforms Mistral OCR 3 on EDGAR-OCR, but Mistral is chosen as the Pareto-optimal production option due to latency.
• Scores on EDGAR-Forecast correlate positively with visible company filing-history size for all four evaluated models.

Limitations

• Does not parse embedded images.
• PDF-derived content depends on OCR quality and parser coverage may lag new or rare XML schema branches.
• Extremely layout-heavy filings yield lower-fidelity reconstructions than standard formats.

Experimental setup

Datasets: EDGAR full archive (1994-present) SEFD-v1

Benchmarks: EDGAR-OCR EDGAR-Forecast HTML table reconstruction benchmark

Metrics: adjusted recall weighted recall exact shape relative-error tolerance score median latency

Baselines: EdgarTools to_markdown Mistral OCR 3 GPT-5.4 Gemini 3.1 Pro Claude Opus 4.7

Builds on: BeanCounter OmniDocBench EdgarTools FinanceAgent

Closest papers by meaning

• FinBen: A Holistic Financial Benchmark for Large Language Models (NeurIPS 2024 · similarity 0.86)
• BeanCounter: A low-toxicity, large-scale, and open dataset of business-oriented text (NeurIPS 2024 · similarity 0.85)
• SILO Language Models: Isolating Legal Risk In a Nonparametric Datastore (ICLR 2024 · similarity 0.83)
• FinSearchComp: Towards a Realistic, Expert-Level Evaluation of Financial Search and Reasoning (ICLR 2026 · similarity 0.83)
• FinMMDocR: Benchmarking Financial Multimodal Reasoning with Scenario Awareness, Document Understanding, and Multi-Step Computation (AAAI 2026 · similarity 0.83)
• What's In My Big Data? (ICLR 2024 · similarity 0.83)