zorynthiq/zoryntiq-sec-filings

Zoryntiq SEC Filings Dataset

A clean, LLM-ready dataset of SEC EDGAR filings from recently IPO'd and pre-IPO companies.

Dataset Summary

5,179 text chunks extracted from 261 high-signal SEC filings (S-1, 10-K, 10-Q, 8-K, DRS, and more). Each chunk is ~1,500 words with 150-word overlap, cleaned and normalized for LLM training and financial NLP tasks.

What's inside

• Registration statements (S-1, S-1/A, DRS) — full IPO prospectuses including business descriptions, risk factors, financials, and use of proceeds
• Annual & quarterly reports (10-K, 10-Q) — MD&A, financial statements, risk disclosures
• Current reports (8-K) — material events, press releases, earnings announcements
• Proxy statements (DEF 14A) — governance, executive compensation
• SEC correspondence (CORRESP) — SEC review letters and company responses

Data fields

Field	Type	Description
id	string	Stable unique ID (SHA-256 of accession + filename + chunk index)
source	string	Always "sec_edgar"
filing_type	string	SEC form type (S-1, 10-K, 8-K, …)
document_type	string	Specific document or exhibit type
document_desc	string	Human-readable document description
company_name	string	Full legal company name
ticker	string or null	Stock ticker symbol
cik	string	SEC Central Index Key
date_filed	string	Filing date (YYYY-MM-DD)
exchange	string or null	Exchange (NASDAQ, NYSE, …)
sic	string	SIC industry code
sic_description	string	SIC industry description
accession_number	string	SEC accession number
url	string	Direct URL to the source document on SEC EDGAR
chunk_index	int	0-based chunk position within the document
total_chunks	int	Total chunks in this document
word_count	int	Word count of this chunk
char_count	int	Character count of this chunk
text	string	Clean plain text (LLM-ready)
financials	dict or null	Key XBRL financial metrics (primary doc chunk 0 only)

financials sub-fields (when present)

Field	Description
revenues_usd	Total revenues
net_income_loss_usd	Net income or loss
assets_usd	Total assets
liabilities_usd	Total liabilities
equity_usd	Stockholders' equity
cash_usd	Cash and cash equivalents
operating_income_usd	Operating income or loss
gross_profit_usd	Gross profit
eps_basic	Basic earnings per share
eps_diluted	Diluted earnings per share
shares_outstanding	Common shares outstanding
period_end	Reporting period end date (YYYY-MM-DD)

Cleaning pipeline

1. Form-type filter — high-signal forms only (S-1, 10-K, 10-Q, 8-K, DEF 14A, DRS, CORRESP, …)
2. Complete-submission filter — skip the monolithic EDGAR .txt wrapper file
3. Exhibit filter — skip EX-31/32/23 boilerplate certifications and binary graphics
4. SGML envelope strip — remove <DOCUMENT>…<TEXT> wrapper
5. HTML → plain text — tables preserved as tab-separated rows
6. Normalisation — control chars, junk lines, page numbers removed
7. Min-word threshold — chunks under 200 words dropped
8. Language filter — English only
9. Content deduplication — MD5 hash; exact duplicates removed
10. Chunking — paragraph-aware ~1,500 words with 150-word overlap

Usage

from datasets import load_dataset

ds = load_dataset("zoryntiq/zoryntiq-sec-filings")

# All S-1 registration statement chunks
s1 = ds["train"].filter(lambda x: x["filing_type"] in ["S-1", "S-1/A"])

# Only chunks with financial data
with_financials = ds["train"].filter(lambda x: x["financials"] is not None)

Source

All data sourced from SEC EDGAR, the U.S. Securities and Exchange Commission's public filing system. EDGAR data is in the public domain.

Citation

@dataset{zoryntiq_sec_filings_2026,
  title   = {Zoryntiq SEC Filings},
  author  = {Zoryntiq},
  year    = {2026},
  url     = {https://huggingface.co/datasets/zoryntiq/zoryntiq-sec-filings},
  license = {Apache 2.0}
}