kurry/amzn_sec_db

AMZN SEC Filings – Chunk-level Corpus (10-K, 10-Q, 8-K)

A ready-to-use, chunk-level corpus of Amazon's (AMZN) recent SEC filings (10-K, 10-Q, and 8-K).
Each paragraph and sentence is stored together with rich metadata, making the dataset ideal for:

• semantic search / RAG pipelines (ChromaDB, FAISS, Weaviate, …)
• question-answering over financial filings
• experimenting with financial-domain embeddings

Time span : last 5 fiscal years (rolling window, as of 2025-05-11)
Collection : 10-K, 10-Q, 8-K (including MD&A summaries & optional exhibits)
Granularity: ~1000-char paragraphs and ≤80-token sentences

---

Contents

1. Dataset structure
2. Download & usage
3. Creation process
4. Intended use
5. Limitations & ethical considerations
6. Citation
7. License
8. Contact

---

Dataset structure

The data are stored in Arrow (the native format of 🤗 datasets) so they can be accessed in-memory or streamed on-the-fly.

Column	Type	Description
text	str	Plain-text chunk (paragraph or sentence)
metadata	dict	All fields defined in FilingMetadata (ticker, cik, filing_type, …)
id	str	SHA-1 hash → unique, deterministic identifier
chunk_type*	str	Paragraph / sentence / summary / exhibit / press_release

*chunk_type is embedded inside metadata.

Total size: ≈ 200-400 k chunks (depends on new filings).

Example metadata object

{
  "ticker": "AMZN",
  "cik": "0001018724",
  "company_name": "AMAZON.COM, INC.",
  "filing_type": "10-K",
  "filing_date": "2025-02-02",
  "filing_period": "2024-12-31",
  "filing_url": "https://www.sec.gov/Archives/...",
  "section_id": "item7",
  "section_title": "Management’s Discussion and Analysis",
  "section_level": 1,
  "chunk_index": 3,
  "chunk_count": 42,
  "chunk_type": "paragraph"
}