Source files are SEC .htm filings submitted between Jan 01, 2000 and Jan 01, 2026.
Files were downloaded on a best-effort basis via EDGAR APIs; some filings may be missing due to source corruption or transient download failures.
In total, 651,900 filings were downloaded and processed.

Background

SEC 10-Q, 10-K, and 8-K filings are rich, high-value disclosures, but the raw HTML is optimized for human page layout rather than machine consumption.

Using raw EDGAR HTML directly for LLM training or inference is inefficient because:

presentation markup (styles, layout wrappers, inline artifacts) can dominate token count without adding meaning,
pagination artifacts (repeated headers/footers, page numbers) interrupt continuity,
textual content is often fragmented by layout-oriented HTML structures.

Intended Uses

We transform raw EDGAR HTML into a stable, model-friendly representation suitable for:

LLM pretraining
long-context modeling
information retrieval / RAG

How

In particular, we:

Remove pagination artifacts: page numbers, repeated headers/footers, and similar print-format remnants.
Normalize text: whitespace, broken line wraps, and inconsistent line breaks.
Preserve structure: parse filings into sections, tables, and figures, then render to consistent text while tracking metadata (e.g., numTables, numFig).
Keep table boundaries explicit using \<|start_table|\> and \<|end_table|\> markers. Tables remain in HTML format because many contain merged cells that are not represented cleanly in Markdown.

Notes

Token counts vary by tokenizer. Counts reported here were computed with the marin-tokenizer and should be treated as approximate length estimates.

Version logs

2025-02-25: [Bug] Fixed an issue where roman page numbers are not detected. [Feature] A single table that was broken up into multiple ones across pages are now merged back into a single table.

Downloads last month: 13