---
language:
- en
license: other
library_name: mlx
pipeline_tag: text-generation
base_model: mlx-community/Qwen2.5-7B-Instruct-4bit
tags:
- sba
- "504"
- "7a"
- lora
- screening
---

# Trace

Trace is a specialized MLX LoRA adapter trained on high-quality, curated synthetic data for SBA 504 and 7(a) screening.

It helps a lender identify the rule that applies on a loan’s effective date, surface a potential hard stop, and prepare a reviewable screening summary. A qualified human must confirm every eligibility and credit decision.

## What is included

- LoRA adapter weights in MLX format
- `adapter_config.json`
- Evaluation visuals based on synthetic screening scenarios

The complete Trace system also uses a dated retrieval layer and deterministic policy checks. Those components are separate from this adapter.

## Training and grounding

The adapter was trained on 11,196 curated synthetic examples and 768 screening stems. The examples use fictional entities and are grounded in these authoritative sources:

- [SBA SOP 50 10 — Lender and Development Company Loan Programs](https://www.sba.gov/document/sop-50-10-lender-development-company-loan-programs)
- [13 CFR Part 120 — Business Loans](https://www.ecfr.gov/current/title-13/chapter-I/part-120)
- [13 CFR Part 121 — Small Business Size Regulations](https://www.ecfr.gov/current/title-13/chapter-I/part-121)
- [13 CFR 120.120 — Eligible use of proceeds](https://www.ecfr.gov/current/title-13/section-120.120)
- [13 CFR 120.131 — Leasing part of an existing or proposed building](https://www.ecfr.gov/current/title-13/section-120.131)

The effective loan date selects the applicable SOP version. The complete system retrieves the source text before producing a screening summary.

## Model details

| Field | Value |
| --- | --- |
| Architecture | Causal language model with LoRA adapter |
| Base model | [Qwen2.5-7B-Instruct 4-bit for MLX](https://huggingface.co/mlx-community/Qwen2.5-7B-Instruct-4bit) |
| Fine-tuning method | MLX LoRA, rank 16, prompt masking enabled |
| Selected checkpoint | v9-160 |
| Primary output | Screening JSON and officer-facing draft summary |
| Intended users | Trained lending and compliance staff |

## Evaluation

ScreenBench holdout uses 12 frozen synthetic files and a deterministic grader.

| Configuration | All-pass | False-kill |
| --- | ---: | ---: |
| Adapter without the Trace runtime | 6/12 | 0 |
| Complete system with retrieval and policy checks | 12/12 | 0 |

These results measure the complete screening workflow on a small internal holdout. They are not an independent certification.

## Quickstart

```python
from mlx_lm import load, generate

model, tokenizer = load(
    "mlx-community/Qwen2.5-7B-Instruct-4bit",
    adapter_path="AmPac/trace",
)
```

The adapter alone does not provide the dated retrieval and deterministic policy checks used by the complete Trace system.

## Important limitations

Trace does not determine eligibility, approve credit, or speak for the U.S. Small Business Administration. It produces review drafts for qualified staff.

Do not use Trace as the sole basis for an approval, decline, servicing action, or applicant communication. Confirm the source text and effective date before relying on an output.

## License and attribution

Repository license: `other`. The base model has its own license and terms. SBA SOPs, federal regulations, and notices remain subject to their source terms.