Dragon Finance US
Qwen2.5-7B-Instruct with a LoRA adapter (r=16, alpha=32) trained on instruction data distilled from the Dragon Finance Corpus — a US-only corpus of public financial data.
Training data (hub-measured)
| metric | value |
|---|---|
| corpus stored | 254.92 GB |
| documents indexed | 117,638 |
| companies covered | 487 unique SEC filers (crawl in progress) |
| filing years | 1999–2026 |
| EDGAR filings | 98,722 — 8-K, 10-Q, 10-K, DEF 14A, S-1, 13F and amendments |
| EDGAR complete submissions | 18,524 filings incl. every exhibit |
| HMDA mortgage records | 392 files, 56 states, 2018–2024 |
Sources
US federal and official only: SEC (filings, bulk data sets, full-index, EDGAR log files, litigation releases), US Treasury (yield curve, fiscal data, auctions), Federal Reserve (Z.1, H.10, SLOOS, DFA, SCF, speeches), FDIC, NCUA, Chicago Fed, CFPB (HMDA, complaints), CFTC, DOL Form 5500, IRS (SOI, 990), SBA, FEMA, USDA NASS, EIA, Cboe, HUD, FHFA, Ginnie Mae, OFAC, Federal Register, BLS, Census, BEA, FRED.
Pipeline
ingest -> corpus -> teacher distillation -> tokenize -> LoRA SFT -> evaluate -> publish,
run end to end by an AI agent in a marimo notebook. Data is stored as tar shards with
sha256 provenance; nothing is fetched from gated, licensed or registration-only sources.
Intended use and limits
Research and informational use. Not investment advice; not an offer or solicitation. Figures are restated by issuers and may be superseded. The corpus is a snapshot and grows; the coverage numbers above will change.