sba-extract-v4 — SBA Form 413 Field Extractor
A fine-tuned Qwen2.5-3B model that extracts 16 financial fields from US tax returns into structured JSON for SBA Form 413 (Personal Financial Statement). Built for community banks and credit unions automating small business lending.
What it does
Feed it a tax return (text or markdown) → get structured JSON:
{
"name": "John M Smith",
"ssn": "123-45-6789",
"employer": "ACME Technologies Inc",
"annual_income": 125000,
"cash": 120000,
"real_estate": 650000,
"mortgages": 450000,
"stocks_bonds": 120000,
"retirement_accounts": 250000,
"life_insurance": 35000,
"notes_receivable": 10000,
"credit_card_debt": 8500,
"auto_loans": 25000,
"total_assets": 1185000,
"total_liabilities": 483500,
"net_worth": 701500
}
Accuracy
| Metric | Score |
|---|---|
| Raw field accuracy | 81% (13/16 exact) |
| With post-processing | 100% (16/16, includes multi-value sum + regex totals) |
| Name/SSN/Employer | 100% |
| Income/Asset extraction | 78-85% confidence per field |
The 3 partial misses are multi-value fields (e.g., two bank accounts, two properties) — the model picks the first value. The Forms SBA platform fixes this with a post-processing sum-of-parts rule.
Model details
| Parameter | Value |
|---|---|
| Base architecture | Qwen2 |
| Parameter count | 3.1B (3,085,938,688) |
| Context length | 32,768 tokens |
| Quantization | F16 (full precision, zero loss) |
| Format | GGUF v2 |
| Attention | GQA (16 query heads, 2 KV heads) |
| Layers | 36 |
| File size | 6.2 GB |
Quick start
Option 1: Ollama
ollama pull hf.co/newdim001/sba-extract-v4
ollama run sba-extract-v4
Option 2: llama.cpp
llama-cli \
-m sba-extract-v4.gguf \
-p "Extract SBA Form 413 fields from: FORM 1040 - John Smith SSN 123-45-6789 Wages: 125000 Cash: 45000..."
Option 3: Python (recommended prompt)
prompt = """
### Instruction:
Extract ALL SBA Form 413 fields from the tax return below. Return ONLY valid JSON with these 16 fields:
- name, ssn, employer, annual_income, cash, real_estate, mortgages
- stocks_bonds, retirement_accounts, life_insurance, notes_receivable
- credit_card_debt, auto_loans, total_assets, total_liabilities, net_worth
### Input:
{your_tax_return_text}
### Response:
"""
Built by
Koodo Technology LLC FZ (Dubai) — part of the Forms SBA platform. Automated SBA lending for US community banks and credit unions.
License
MIT