fin-ai-lab/BeanCounter-v2

BeanCounter v2 (beta)

Faithful, typed extraction of every document disseminated on SEC EDGAR, 1996 → 2026-08: 32,851,820 document records across 18,421,669 filings (all firms, all form types), decomposed into 4,874,403,057 ordered content blocks (headings, paragraphs, tables) that reproduce what a human sees in the browser. Successor to BeanCounter (paper).

Beta notes: this release ships the index and blocks tables. Two document-level payloads are extracted but not yet published — per-form structured data (has_data, e.g. every Form 4 transaction / 13F holding / NPORT-P position as typed fields) and byte-exact EDGAR-rendered HTML (has_rendered_html). They will arrive as an additional config. Image binaries are not included; image positions are preserved as image_ref blocks.

Keys, dates, and re-disseminations

One filing (accession) contains many documents: the main body, each exhibit, each structured XML doc, each OCR'd PDF. EDGAR occasionally re-disseminates the same filing (mostly <CORRECTION> events, common before ~2001), so the same (accession, filename) can appear on several days — sometimes with differing content (the corrected copy's feed body is often truncated by EDGAR; occasionally it is a genuine content replacement). This dataset preserves every disseminated copy and takes no editorial position on which is "the" document.

The unique key is (accession, filename, dissemination_date, dissem_seq):

• filing_date — EDGAR's <FILING-DATE> from that copy's own header.
• dissemination_date — the day the copy went out on the public feed (this also drives the year=/month= partitioning).
• dissem_seq — ordinal among same-key copies within one day (0 for 99.97% of records).

To find re-disseminated documents: GROUP BY accession, filename HAVING count(*) > 1. To get one copy per document, a common choice is the largest text_chars per (accession, filename) (the feed's corrected copies are usually truncated), or the earliest/latest dissemination_date depending on your research design.

Tables

index — one row per document copy (32,851,820 rows)

column	type	meaning
accession	string	filing accession number (dashed)
filename	string	document filename within the filing
dissemination_date	date32	feed day of this copy (key part)
dissem_seq	int32	same-day copy ordinal (key part, usually 0)
extension	string	filename extension
form_type	string	the FILING's form (e.g. 8-K)
doc_type	string?	this DOCUMENT's <TYPE> (e.g. EX-99.1)
sequence	string?	document sequence within the submission
doc_format	string?	extraction route: html | text | xml_struct | pdf
filing_date	date32	EDGAR filing date from this copy's header
cik	int64?	primary filer CIK
ciks	list	all filing entities {cik, role, name} (issuer, reporting_owner, filer, ...)
n_blocks	int32	number of content blocks
n_tables	int32	number of table blocks
text_chars	int64	total characters of block text
n_warnings	int32	parser warnings count
has_error	bool	a per-document extraction error was recorded
has_data	bool	typed structured payload exists (future config)
has_rendered_html	bool	byte-exact EDGAR HTML exists (future config)

blocks — one row per content block (4,874,403,057 rows)

Key columns plus duplicated hot-filter columns so most queries never touch the index.

column	type	meaning
accession, filename, dissemination_date, dissem_seq		document key (join to index)
filing_date, form_type, doc_type, doc_format, cik		duplicated filter columns
block_order	int32	position within the document
block_type	string	heading | text | table | image_ref
text	string?	heading/text content
heading_level	int32?	1–6 for headings
page	int32?	logical page (filer page-break markers; olmOCR pages for PDFs)
table_grid	list<list>?	canonical lossless cell matrix (exact source cell values)
table_markdown	string?	rendered GFM view (may drop empty spacer columns)
table_n_rows, table_n_cols	int32?	grid dimensions
table_is_complex	bool?	merged cells / multi-row headers present

Reconstructing a document as markdown: select its blocks ordered by block_order, render headings as '#' * heading_level + ' ' + text, tables via table_markdown, and join with \n\n:

def to_markdown(blocks):
    parts = []
    for b in sorted(blocks, key=lambda b: b["block_order"]):
        if b["block_type"] == "heading":
            parts.append("#" * (b["heading_level"] or 1) + " " + (b["text"] or ""))
        elif b["block_type"] == "table":
            parts.append(b["table_markdown"] or "")
        elif b["text"]:
            parts.append(b["text"])
    return "\n\n".join(parts)

Physical layout

data/index/year=YYYY/part-0000.parquet            31 files (1996–2026)
data/blocks/year=YYYY/month=MM/part-NNNN.parquet  zstd, ~≤600 MB parts
manifests/shards.parquet                          per-file rows, dates, bytes, sha256

Partitions follow dissemination_date. Rows are ordered by dissemination date, so parquet column statistics prune date-range queries even without hive-partition filtering. Use manifests/shards.parquet to plan partial downloads (it lists every data file's row count, date span, size, and sha256).

Usage

Stream blocks with datasets

from datasets import load_dataset

blocks = load_dataset(
    "fin-ai-lab/BeanCounter-v2", "blocks", split="train", streaming=True,
    columns=["accession", "filename", "dissemination_date", "form_type",
             "block_order", "block_type", "text", "table_markdown"],
)
for b in blocks.take(20):
    print(b["form_type"], b["block_type"], (b["text"] or "")[:80])

Query the index with DuckDB

SELECT filing_date, form_type, count(*) AS documents
FROM read_parquet('hf://datasets/fin-ai-lab/BeanCounter-v2/data/index/**/*.parquet')
WHERE cik = 320193
GROUP BY ALL ORDER BY filing_date;

Fetch one filing's blocks (DuckDB, pruned by date)

SELECT * FROM read_parquet(
  'hf://datasets/fin-ai-lab/BeanCounter-v2/data/blocks/year=2024/month=08/*.parquet')
WHERE accession = '0000320193-24-000081'
ORDER BY filename, dissem_seq, block_order;

Link back to EDGAR

def edgar_filing_url(accession, cik):
    return (f"https://www.sec.gov/Archives/edgar/data/{cik}/"
            f"{accession.replace('-', '')}/{accession}-index.htm")

Provenance & method

• Source: complete EDGAR dissemination feed (daily oldloads), 1996 → 2026-08-03, all firms and all form types; 433,084 PDF documents OCR'd with olmOCR; 7,885,083 structured XML documents parsed by per-form deterministic parsers (validated byte-exact against EDGAR's own rendering); HTML and text documents extracted by engines validated against vision-model ground truth on the 109 highest-volume HTML form types.
• Coverage: 100%+ of filings listed in EDGAR's own indexes for the period (the dissemination feed is a superset, including correction re-disseminations preserved here). SEC "paper" placeholder accessions (9999999997-*) have no electronic content and are absent. The only dropped records are 2 byte-identical extraction-pipeline artifacts.
• Tables: table_grid is the lossless representation; table_markdown is a presentation view. 210 pathological tables (>2M cells, phantom-column colspan artifacts and dense fund-vote dumps) retain full grids but skip a cosmetic column-merge pass.

Citation

If you use BeanCounter v2, please cite:

@inproceedings{wang2024beancounter,
  title={BeanCounter: A low-toxicity, large-scale, and open dataset of business-oriented text},
  author={Wang, Siyan and Levy, Bradford},
  booktitle={Advances in Neural Information Processing Systems},
  year={2024}
}