Dragon Finance US
Qwen2.5-7B-Instruct, LoRA-fine-tuned (r=16) by knowledge distillation from Qwen2.5-7B-Instruct teacher-generated samples over official US financial data.
Companion dataset repo: datasets/dragonlimited/Dragon-Finance-US
(raw data, chunked corpus, tokenized SFT data, teacher distill samples).
Data domains: SEC/EDGAR filings (10-K, 10-Q, 8-K, S-1, 13F, DEF 14A) + XBRL financials, FINRA rules, NYSE (newsroom/rules), NASDAQ daily files, EIA commodities, CFTC futures COT (incl. crypto futures), US Treasury yield curve, FRED rates, crypto spot (commercial API, flagged non-official).
Pipeline: ingest -> chunk -> teacher distillation (extract/qa/explain/reconcile) -> tokenize (with holdout split) -> LoRA SFT -> eval on holdout -> upload. Full spec lives in the project marimo notebook.
Layout (this repo)
checkpoints/final/LoRA adaptercheckpoints/merged/merged base + adapter weightsREADME.mdthis model card
License notes: base model = Qwen2.5 (Qwen license). Training data includes US government public records (SEC, Treasury, EIA, CFTC), CC-BY-4.0 (FRED), and free official exchange content. NYSE/NASDAQ market tape and FINRA bulk data are NOT included (paid/licensed).