astr010/sec-10k-lsh-chunks

📈 SEC 10-K Cleaned Text Chunks & LSH Boilerplate Dataset

Dataset Summary

This dataset contains 13,562,130 cleaned text chunks extracted from 12,361 SEC Form 10-K annual filings across 1,380 companies (spanning 2004 to 2025, core 2014–2025).

Every chunk across all 1,380 companies (including mega-cap leaders such as AAPL, MSFT, NVDA, AMZN, GOOGL, META, TSLA, JPM, WMT, XOM, AVGO, LLY) is annotated with metadata, token counts, table indicators, and a pre-computed MinHash Locality-Sensitive Hashing (LSH) boilerplate classification label (is_boilerplate) to identify repetitive legal disclosure boilerplate across filings.

---

📊 Dataset Statistics

Metric	Value
Total Companies (Tickers)	1,380 companies (100% LSH annotated)
Total 10-K Filings Processed	12,361 filings
Total Text Chunks	13,562,130 chunks
File Format	Parquet (1 file per ticker)
Fiscal Years Covered	2004 – 2025 (Core: 2014–2025)
Primary Language	English
Licensing	MIT

---

🧩 Schema & Features

Each row represents a ~300–500 token text chunk from an SEC Form 10-K filing:

Field Name	Type	Description
ticker	string	Stock ticker symbol (e.g., AAPL, MSFT, AMZN, NVDA)
cik	string	SEC Central Index Key (10-digit zero-padded ID)
sic	string	Standard Industrial Classification code (e.g., 3571)
sic_description	string	Industry classification text (e.g., Electronic Computers)
fiscal_year	int64	Fiscal year of the 10-K filing (e.g., 2024)
chunk_id	int64	Sequential chunk index within the filing document
is_table	bool	True if the chunk contains tabular data, False for narrative text
text	string	Cleaned text content of the chunk
token_count	int64	Number of tokens in the chunk
is_boilerplate	bool	True if flagged as repetitive legal/financial boilerplate via MinHash LSH

---

💻 Quickstart & Usage

Using Hugging Face datasets:

from datasets import load_dataset

# Load dataset
dataset = load_dataset("astr010/sec-10k-lsh-chunks")

# Sample a chunk
sample = dataset["train"][0]
print(f"Ticker: {sample['ticker']} ({sample['fiscal_year']})")
print(f"Is Boilerplate: {sample['is_boilerplate']}")
print(f"Text Preview:\n{sample['text'][:200]}...")

Loading directly with pandas / pyarrow:

import pandas as pd

# Load chunks for a specific ticker directly from the repository
df = pd.read_parquet("hf://datasets/astr010/sec-10k-lsh-chunks/AAPL.parquet")
print(df.head())

---

🛠️ Data Pipeline & Methodology

1. Extraction: Raw SEC 10-K HTML filings downloaded via SEC EDGAR.
2. HTML-to-Markdown Conversion: Standardized into clean Markdown format preserving headers and Markdown tables.
3. Chunking: Chunked into natural text blocks (~300–500 tokens), separating narrative paragraphs from financial tables (is_table).
4. Boilerplate Detection (MinHash LSH): Evaluated using a bounded-memory streaming MinHash LSH pipeline to detect near-duplicate legal disclosures and recurring regulatory boilerplate across filings and companies.

---

📜 Licensing & Citation

This dataset is released under the MIT License.

@dataset{sec_10k_lsh_chunks_2026,
  author       = {Tim Condello},
  title        = {SEC 10-K Cleaned Text Chunks with MinHash LSH Boilerplate Annotations},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/astr010/sec-10k-lsh-chunks}}
}