CoolFace
Datasetpublic

CyberMax-tools/riskroll-sec-10k-10q-sections

Riskroll: SEC 10-K and 10-Q sections as clean text Try the paid tool: Boardroom Buys on Apify: track insider buys at the companies behind these reports. First try costs cents; pay only for results. The parts of annual and quarterly reports that analysts, researchers and LLM pipelines actually read, cut out of each filing and cleaned: Risk Factors (Item 1A), Management's Discussion and Analysis (MD&A), Quantitative and Qualitative Disclosures About Market Risk, Cybersecurity… See the full description on the dataset page: https://huggingface.co/datasets/CyberMax-tools/riskroll-sec-10k-10q-sections.

sourceHugging Facecc0-1.0updated 8h agoView on Hugging Face
0likes
Dataset Card

Riskroll: SEC 10-K and 10-Q sections as clean text

<!-- cybermax-more:start -->

Try the paid tool: Boardroom Buys on Apify: track insider buys at the companies behind these reports. First try costs cents; pay only for results.

<!-- cybermax-more:end -->

The parts of annual and quarterly reports that analysts, researchers and LLM pipelines actually read, cut out of each filing and cleaned: Risk Factors (Item 1A), Management's Discussion and Analysis (MD&A), Quantitative and Qualitative Disclosures About Market Risk, Cybersecurity (Item 1C) and Business (Item 1), one row per filing × section, for every 10-K and 10-Q filed with the SEC. New filings are added as they come in, so the corpus grows month by month.

This build (see meta.json): 2,098 filings (10-K, 10-Q and amendments) filed 2026-08-12 to 2026-09-23 (plus one filing dated 2026-02-23 that EDGAR indexed in September; 1,858 10-Q, 170 10-K, 67 amendments, 3 transition reports; the tail of the Q2 10-Q season plus fiscal-June annual reports), 5,245 sections: 1,886 MD&A, 1,560 Risk Factors, 1,441 Market Risk, 185 Business and 173 Cybersecurity, 20.4 million words in total. 1,979 of the filings have at least one section (Part III-only amendments and exhibit-only filings have none, by design; 63 filings from the 12–14 Aug peak couldn't be fetched because SEC was throttling and are marked sec_fetch_failed; see filings).

Why this and not the popular EDGAR corpora? The best-known 10-K section corpus on Hugging Face (EDGAR-CORPUS) was last updated in July 2023 and has annual reports only; full-filing dumps leave the section splitting to you. This one starts with 2026 filings, includes 10-Qs (quarterly MD&A and risk-factor updates) and the new Item 1C cybersecurity disclosures, and keeps adding filings.

Use it

python
from datasets import load_dataset
ds = load_dataset("CyberMax-tools/riskroll-sec-10k-10q-sections", "sections", split="train")

# Every risk-factor section that mentions tariffs
risky = ds.filter(lambda r: r["section"] == "risk_factors" and "tariff" in r["text"].lower())
print(len(risky), risky[0]["company_name"], risky[0]["filing_url"])

DuckDB, straight from the Hub:

sql
SELECT ticker, company_name, form_type, filed_date, n_words
FROM 'hf://datasets/CyberMax-tools/riskroll-sec-10k-10q-sections/data/sections-*.parquet'
WHERE section = 'cybersecurity' ORDER BY n_words DESC LIMIT 20;

Good for: RAG over company filings, risk-factor change tracking (compare a company's 10-K with its next 10-Qs), classifier and summarizer training/evaluation, sector-wide risk themes (AI, tariffs, cyber incidents, rates), and giving an AI agent the "why" behind the numbers.

Configs and columns

`sections` (default, one row per filing × section): accession_number, form_type, filed_date, period_of_report, cik, ticker, company_name, section (risk_factors, mdna, market_risk, cybersecurity, business), item (1A, 7, 7A, 1C, 1 for 10-K; II-1A, I-2, I-3 for 10-Q), text, n_chars, n_words, filing_url (EDGAR index page), document_url (the primary document the text came from).

`filings` (one row per 10-K/10-Q filing in the window): the same identifiers plus sections_found and status (ok, no_sections_found, no_html_primary_document, sec_fetch_failed), so you can see exactly what was and wasn't extracted.

How the text is made: the filing's primary HTML document is converted to text line by line (tables kept as cell | cell rows, page numbers and "Table of Contents" lines dropped). A section runs from its Item heading to the next Item heading; when a heading appears more than once (table of contents, cross-references), the occurrence with the longest body is used. Nothing is paraphrased or summarized. Smaller reporting companies often write "not required" under Market Risk; those short sections are kept as filed. Parsing is heuristic, so a small share of filings with unusual layouts can have a section missing or cut early; check document_url when it matters.

Refresh

New filings are added weekly from the SEC's daily indexes; files are per filing month (data/sections-YYYY-MM.parquet). meta.json has the latest build's window and counts.

Licence

SEC EDGAR filings are US government public records, free of copyright restrictions; this compilation is released under CC0 1.0. Compiled by CyberMax. Not affiliated with or endorsed by the SEC. Not investment advice.

Also from CyberMax

<!-- cybermax-xlinks:start -->

More free from CyberMax