dragonlimited/Dragon-Finance-US

Dragon Finance US — US Government Open Finance Dataset

Collection of official, directly-downloadable US-finance data gathered from primary federal agencies, cleaned, chunked, and pre-tokenized to distill a Qwen2.5-7B-Instruct finance model. Data artifacts live in this repo; the fine-tuned weights will live in a separate model account (HF_MODEL_NAMESPACE).

Scope rule: only sources that are keyless, form-free, and directly downloadable are kept. Sources requiring API keys, interactive forms, or paid access are intentionally excluded.

⚠️ IMPORTANT — READ BEFORE USE. This dataset is provided for research and informational purposes only and is not investment, financial, tax, legal, or accounting advice, and not an offer or solicitation to buy or sell any security or instrument. See the full Legal Disclaimer at the end of this card.

✅ Done

• Repos created: datasets/dragonlimited/Dragon-Finance-US (data) and dragonlimited/Dragon-Finance-US (model / LoRA placeholder).
• End-to-end pipeline validated (smoke): ingest -> chunk -> teacher distillation (Qwen2.5-7B-Instruct) -> tokenize -> LoRA SFT -> eval -> upload.
• Implemented keyless source pipelines: SEC EDGAR filings, SEC XBRL companyfacts, SEC Financial Statement Data Sets (FTD full universe), US Treasury yields, Treasury Fiscal, FRED, CFTC COT, FINRA rules, Nasdaq/NYSE free feeds, BLS CPI/PPI, FDIC institutions, CFPB complaints, HMDA (state-by-state), NCUA call reports.
• Parallel multi-download engine (MAX_WORKERS thread pool), resumable / idempotent downloads, sha256 provenance manifest, auto-namespaced HF uploads.
• Auto-updating card + progress chart (this README.md + progress.png).

🚧 Doing (in progress)

• Filling the local volume with full-history US-government data.
• SEC EDGAR full universe (all US filers) + full filing history.
• SEC FTD full universe 2009q1 -> 2026.
• HMDA all states x 2007-2024; NCUA call reports 2000 -> 2026.
• Treasury 1990 -> today; CFTC COT 1986 -> today.
• FDIC ~100k institution records; CFPB complaints bulk.
• All direct sources downloading concurrently (16 workers + keep-alive sessions).

🔜 Upcoming (pipeline scale + quality)

• Tokenizer-native chunking (replace word-window approximation).
• Full-corpus teacher distillation: extract / qa / explain / reconcile.
• vLLM teacher batching for speed.
• Full LoRA SFT, holdout benchmark, merge + quantize, publish to the separate model account.
• Quality loop: dedup, numeric-fact QA, unit reconciliation, iterative filtering.

Directory layout

data/        raw downloads by source folder + manifest.jsonl (sha256)
corpus/      chunks.jsonl  (cleaned text chunks)
distill/     train.jsonl   (teacher-generated instruction/answer samples)
tokenized/   sft/          (HuggingFace Arrow dataset) + info.json

Data sources (all direct / keyless)

domain	agency	dataset	status
disclosures	SEC	EDGAR 10-K/10-Q/8-K/S-1/13F/DEF 14A	✅ live
financial statements	SEC	XBRL companyfacts	✅ live
XBRL all US filers	SEC FTD	sub / num / pre, 2009q1->now	✅ live
rates / bonds	US Treasury	daily yield curve	✅ live
fiscal rates	US Treasury Fiscal Data	avg interest rates	✅ live
macro rates	FRED	10Y / 30Y / fed funds / CPI	✅ live
futures	CFTC	COT compiled zip	✅ live
broker rules	FINRA	rulebook	✅ live
exchange free feeds	Nasdaq / NYSE	OI / volume, newsroom	✅ live
inflation	BLS	CPI / PPI flat files	✅ live
banking	FDIC	institutions API	✅ live
consumer credit	CFPB	complaints bulk	✅ live
mortgage loans	CFPB / FFIEC	HMDA state-by-state	✅ live
credit unions	NCUA	call-report zips	✅ live

🧭 Pipeline

flowchart LR
    A[ingest] --> B[corpus]
    B --> C[distill]
    C --> D[tokenize]
    D --> E[train]
    E --> F[eval]
    F --> G[upload]
    style A fill:#1f77b4,color:#fff
    style B fill:#1f77b4,color:#fff
    style C fill:#1f77b4,color:#fff
    style D fill:#1f77b4,color:#fff
    style E fill:#1f77b4,color:#fff
    style F fill:#1f77b4,color:#fff
    style G fill:#1f77b4,color:#fff

Corpus schema (corpus/chunks.jsonl)

{"text": "...", "meta": {"source": "sec_edgar", "file": "sec_edgar/CIK...", "chunk": 0}}

Distill schema (distill/train.jsonl)

{"instruction": "task prompt + document chunk",
 "output": "teacher answer",
 "task": "qa | explain | extract | reconcile",
 "domain": "source folder",
 "source_file": "relative path to raw file"}

Tokenization scheme (tokenized/)

• Tokenizer: Qwen2.5-7B-Instruct (pad 151643, eos 151645).
• Chat template: user=instruction, assistant=answer; apply_chat_template -> ["input_ids"].
• Label masking: prompt tokens set to -100; loss on assistant tokens only.
• Padding: left-pad ids with pad token; labels -100; attention 0 on padding.
• Truncation: right-truncate to TRAIN_MAX_SEQ_LEN (3072).
• Holdout: deterministic tail 20% (min 2 rows).
• Artifacts: Arrow dataset + tokenized/info.json.

Rebuild

run_ingest -> run_corpus -> run_distill -> run_tokenize -> run_train -> run_eval -> run_upload. Set SMOKE=False and tune FIN7B_WORKERS for concurrency.

⚠️ Legal Disclaimer & Compliance — Please Read Carefully

By accessing or using this dataset (the "Data"), you acknowledge and agree to the terms below. "Sources" means the US federal and official agencies from which the Data is compiled; "we" means the compilers, contributors, and repository operators.

1. No Offer, Solicitation, or Investment Advice

The Data is provided for general informational and research purposes only. Nothing in the Data or in any derivative model, output, notebook, or repository constitutes (i) an offer or solicitation to buy or sell any security, commodity, derivative, currency, or other financial instrument, (ii) investment, financial, tax, legal, regulatory, or accounting advice, or (iii) a suitability assessment, endorsement, or recommendation. Consult your own qualified professional advisers before making any investment or financial decision.

2. No Fiduciary, Advisory, or Client Relationship

Access to or use of the Data does not create any fiduciary, advisory, agency, broker-dealer, investment-management, client, or trust relationship between you and any Source, compiler, contributor, or repository operator. No such party acts as your adviser or owes you any suitability, best-interest, or disclosure duty by virtue of the Data.

3. No Reliance; Independent Verification Required

The Data may be incomplete, delayed, restated, amended, misfiled, or superseded by a Source. You must independently verify all information before relying on it, and you must not rely on the Data as the sole basis for any decision, valuation, or transaction.

4. Third-Party Data, Marks, and No Endorsement

The Data is compiled from publicly available US federal and official sources, including but not limited to the SEC, US Treasury, Federal Reserve/FRED, CFTC, FDIC, CFPB, BLS, NCUA, FINRA, and the free Nasdaq/NYSE feeds. Inclusion does not imply sponsorship, endorsement, approval, or affiliation by any Source. Agency names and marks are used solely to identify the origin of data and remain the property of their respective owners.

5. Licensing, Attribution, and Redistribution Obligations

Third-party data is subject to the licenses, terms, attribution, and rate-limit requirements of each Source. FRED data is licensed under CC-BY-4.0 and requires attribution. SEC EDGAR is a US public record, but the SEC requires a "Name email" User-Agent and compliance with its access policy. HMDA data is subject to FFIEC/CFPB terms. You are solely responsible for reviewing and complying with all applicable third-party terms before any redistribution, commercial use, or derivative publication.

6. No Licensed, Paid, or Proprietary Content

Licensed or proprietary market data (including NYSE/NASDAQ consolidated tape data and FINRA market data), CME historical data, subscription data, and key-gated data are intentionally NOT included. No right, license, or interest in any such excluded data is granted or implied.

7. No Warranty; "As Is" and "As Available"

The Data is provided "as is" and "as available," without any representation, warranty, or condition of any kind, express, implied, or statutory, including, without limitation, merchantability, fitness for a particular purpose, accuracy, completeness, timeliness, title, and non-infringement. Sources correct and revise their data, and the Data may not reflect the latest revisions.

8. Limitation of Liability

To the maximum extent permitted by law, no compiler, contributor, or repository operator shall be liable for any direct, indirect, incidental, special, consequential, exemplary, or punitive damages, or for any loss of profits, revenue, business, goodwill, data, or use, arising out of or relating to your access to, use of, or inability to use the Data, even if advised of the possibility of such damages.

9. Regulatory Compliance, Lawful Use, and Fair Use

You shall use the Data only in compliance with all applicable laws, rules, and regulations (including securities, financial-services, export-control, and sanctions laws), and you shall respect each Source's rate limits, terms of service, User-Agent requirements, and fair-use expectations. You shall not use the Data in any unlawful, deceptive, manipulative, or infringing manner.

10. Privacy and Prohibition on Re-Identification

Certain public microdata (such as consumer complaints and HMDA mortgage records) is released in a privacy-protected form. You shall not attempt to re-identify, de-anonymize, link, or match such records to any natural person, household, or entity beyond what the Source itself discloses.

11. Forward-Looking and Historical Information

Filings, figures, and statements may be historical, may contain forward-looking statements that are subject to risks and uncertainties, and may be superseded or amended at any time. No assurance is given that any past information is predictive of future results or that any forward-looking statement will be realized.

12. Changes and Acceptance

These terms may be updated at any time without notice. Your continued use of the Data after changes are posted constitutes acceptance of the revised terms. If you do not agree with any part of these terms, do not use the Data.

📊 Live dataset progress (snapshot)

category	files	size
sec_financial_statement_sets	426	37.95 GB
hmda	298	37.92 GB
sec_edgar	96,979	33.80 GB
cfpb	2	6.07 GB
ncua	1,083	3.10 GB
bls	2	113.87 MB
cftc	44	71.89 MB
fdic	28	66.94 MB
sec_xbrl	4	14.68 MB
fred	53	4.71 MB
corpus	1	0.95 MB
nasdaq	15	0.64 MB
treasury	37	0.61 MB
tokenized	4	0.17 MB
finra	1	0.09 MB
distill	1	0.03 MB