DragonData-FinDense-3B
The open bilingual (EN/CN) finance expert that beats giants where it matters — and shows its work. 3B dense, 100% self-trained (own architecture config, own tokenizer, own corpus, own weights). Runs on a laptop at Q4 — including a 1.27 GB pocket build.
Status: IN TRAINING (
dragonlimited— one public home for data + models).
Why it is strong
| # | Claim | Status |
|---|---|---|
| 1 | Beats 70B generalists on finance reasoning (Fin-RATE-style longitudinal + cross-entity) | ☐ target |
| 2 | Bilingual EN/CN finance: SEC + HKEX/A-share/PBOC | ☐ target |
| 3 | Exact deal math (digit-aware numbers + RLVR) | ☐ target |
| 4 | Every figure cited to source cell/line (native grounding) | ☐ target |
| 5 | Reads tables + charts natively (text→table→chart training) | ☐ target |
| 6 | Temporal honesty: annual-cutoff DatedDragon variants, no lookahead | ☐ target |
| 7 | Monthly living releases, knowledge-date stamped | ☐ target |
| 8 | SAFR-shaped: typed, bounded, logged actions for agentic finance | ☐ target |
| 9 | Fully auditable: open corpus, registry, code, checkpoints, GPU-hours | ☑ by construction |
Honest scope: reader/reasoner, never predictor. It will not forecast returns — no model can (sentiment ≠ alpha). It reads filings, reconciles periods, cites sources.
How to use
from transformers import AutoTokenizer, AutoModelForCausalLM
tok = AutoModelForCausalLM # placeholder until first release — see Progress
First release will ship: pytorch_model (bf16) + GGUF Q4_K_M / Q5_K_M / Q8_0 +
Q2 pocket build, tokenizer/ (DragonData 64k finance BPE), eval/ (score sheets),
RUN.md (repro recipe: tokens, GPU-hours, seeds).
Where to find things
├── README.md ← you are here (model card)
├── RUN.md ← exact repro: data snapshot, config, seeds, GPU-hours
├── eval/ ← FinanceBench, FinQA, TAT-QA, Fin-RATE-style, CN-FinEval sheets
├── tokenizer/ ← DragonData 64k finance BPE
├── checkpoints/ ← dated releases (living model), knowledge-date stamped
├── dated/ ← DatedDragon annual-cutoff variants
└── gguf/ ← quantized builds incl. 1.27 GB pocket
Data: dragonlimited/DragonData-Finance-Corpus (open corpus + registry).
Code & plan: DragonData project. Evals first: DragonData-CN-FinEval.
License
Apache 2.0 (code + weights). Training data per-document licenses govern slices —
see corpus metadata/source_registry.json.
Progress
S0 eval → S1 tokenizer → S2 1.2B pilot → S3 3B flagship → S4 distill → S5 MoE → S6 SFT/DPO → S7 RLVR → S8 DatedDragon + living loop → S9 quants. Gates in PLAN.md.