A compact, on-device model that extracts and classifies financial transactions from SMS and email into structured JSON. Fine-tuned from Qwen3-1.7B for the WealthWise local-first personal-finance app.

🚧 Status: experimental / under active development. WealthWise's larger 14B model is the most accurate today; this 1.7B is the phone-sized variant that is catching up. On controlled benchmarks it already matches or beats the 14B; on the long tail of real-world mail it is still improving via distillation from the 14B teacher and user corrections. It is ~8× smaller and several× faster and is intended for laptops, low-power machines, and mobile.

What it does

Given a raw SMS or email message, it returns a single JSON object:

json

{
  "amount": 437.0, "currency": "INR", "type": "debit",
  "merchant": "Swiggy - Behrouz Biryani", "account_last4": null,
  "date": "2026-06-12", "payment_method": "UPI", "reference": null,
  "balance": null, "is_recurring": false,
  "category_l1": "Food & Dining", "category_l2": "Food Delivery"
}

category_l1 is one of: Food & Dining, Transportation, Shopping, Housing, Utilities, Entertainment, Health, Education, Travel, Investments, Financial, Income, Miscellaneous. Non-transactional messages return {"is_transaction": false}.

Files

wealthwise-1.7b-q8_0.gguf — Q8_0 quantized weights (~1.8 GB, near-lossless).
Modelfile — Ollama recipe (correct non-thinking Qwen3 template + system prompt).

Run it

Ollama (recommended)

bash

# from this folder:
ollama create wealthwise-1.7b -f Modelfile
ollama run wealthwise-1.7b
# or, if published to the Ollama registry:
# ollama pull codedrivehg/wealthwise-1.7b

Download the GGUF:

hf download codedrivehg/wealthwise-1.7b-GGUF --local-dir wealthwise-1.7b

llama.cpp

llama-cli -m wealthwise-1.7b-q8_0.gguf -p "<your message>"

Important: this model is trained for non-thinking output (direct JSON). The bundled Modelfile pre-fills the empty <think></think> block Qwen3 expects — use it (or replicate that template) or the model may emit garbage.

Training

Base: Qwen/Qwen3-1.7B (Apache-2.0).
Method: LoRA fine-tune (attention + MLP) on financial extraction data, then merged to 16-bit and exported to GGUF.
Improving via knowledge distillation from the WealthWise 14B teacher (and a larger teacher), an expanding curated merchant database, and in-app corrections.

Limitations

Out-of-distribution real-world emails (unusual formats, multilingual, receipts with many line items) are where it still trails the 14B.
Use the 14B for maximum accuracy on capable hardware.

License

Apache-2.0 (inherits the Qwen3-1.7B base license).