mzwin9870 commited on
Commit
6de04a3
·
verified ·
1 Parent(s): b109458

Model card: repo contents, occupancy 120.131, eval table, load.

Browse files
Files changed (1) hide show
  1. README.md +54 -20
README.md CHANGED
@@ -14,49 +14,83 @@ pipeline_tag: text-generation
14
 
15
  # Trace
16
 
17
- **Occupancy opens 13 CFR 120.131.**
18
 
19
- LoRA on Qwen2.5-7B-Instruct for 504/7(a) origination screening. Paste a file. It must open the SOP or CFR paragraph in force on the loan-number date. A designated officer still certifies eligibility and credit.
20
 
21
- Not an SBA product. Not for applicants. Outputs are drafts. `verified: false`.
22
 
23
- Existing buy → [13 CFR 120.131](https://www.ecfr.gov/current/title-13/section-120.131) **(b)** 51%. New construction → **(a)** 60%. Working capital on a 504 → [13 CFR 120.120](https://www.ecfr.gov/current/title-13/section-120.120) **(b)**. Loan number on 2026-10-15 → **SOP 50 10 8.1**.
24
 
25
- ## Load
26
 
27
- These files are an MLX LoRA. They are not the staff runtime by themselves.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
28
 
29
  ```python
30
  from mlx_lm import load, generate
 
31
  model, tokenizer = load(
32
  "mlx-community/Qwen2.5-7B-Instruct-4bit",
33
  adapter_path="AmPac/trace",
34
  )
35
  ```
36
 
37
- Staff path (TRACE binds the board, occupancy still 120.131):
38
 
39
  ```bash
40
- python3 -m sba_screen --talk --as-of 2026-01-15 --program 504
 
41
  ```
42
 
43
- ## ScreenBench (holdout n=12)
44
-
45
- | Seat | All-pass | False-kill |
46
- |---|---|---|
47
- | Naked adapter (this repo) | 6/12 | 0 |
48
- | Complete system (lake + TRACE bind) | 12/12 | 0 |
49
 
50
- Do not quote 12/12 as a naked-7B number. Exam-v1 with bind is 76/76, fk 0.
 
 
51
 
52
- ## Use
53
 
54
- **Intended:** CDFI, CDC, and 7(a) lender staff drafts.
55
 
56
- **Not:** applicants, credit decisions, SBA determinations, loading this adapter as a chatbot.
57
 
58
- SOP/CFR is **retrieved**. It was not continued-pretraining.
59
 
60
  ## License
61
 
62
- `other`. Qwen2.5 (Tongyi Qianwen) + U.S. government SOP/CFR + screening code in the AmPac repo. No live borrower files in train.
 
14
 
15
  # Trace
16
 
17
+ This repository contains LoRA weights and `adapter_config.json` for Trace, in the MLX format used by `mlx_lm`.
18
 
19
+ Trace is a small open-source 504/7(a) screening model for CDFIs and lenders. Ask it a file. It returns the SOP or CFR paragraph in force on the loan-number date, a hard-stop or bounce, and a board you can check. A human certifies eligibility and credit.
20
 
21
+ **Occupancy opens [13 CFR 120.131](https://www.ecfr.gov/current/title-13/section-120.131).**
22
 
23
+ Working capital on a 504 opens [13 CFR 120.120](https://www.ecfr.gov/current/title-13/section-120.120). SOP 50 10 8.1 does not leak onto a January 2026 file.
24
 
25
+ These artifacts load on the public base [`mlx-community/Qwen2.5-7B-Instruct-4bit`](https://huggingface.co/mlx-community/Qwen2.5-7B-Instruct-4bit). The staff runtime is the dated lake plus TRACE bind — not this adapter alone.
26
 
27
+ ## Highlights
28
+
29
+ - **Occupancy.** Investor real estate and low occupancy open 13 CFR 120.131. Existing buy is **(b) 51%**. New construction is **(a) 60%**.
30
+ - **Proceeds.** Payroll, inventory, and marketing on a 504 open 13 CFR 120.120 **(b)** and fail `USE_OF_PROCEEDS`.
31
+ - **Clock.** A loan number on or after 2026-10-01 opens **SOP 50 10 8.1**. January files stay on SOP 50 10 8.
32
+ - **Talk.** Messy processor English. Follow-up (“what if they occupy 100”) keeps the file. TRACE re-runs.
33
+
34
+ If it cannot open the paragraph, it is not a screen.
35
+
36
+ ## Model overview
37
+
38
+ | | |
39
+ |---|---|
40
+ | Type | Causal LM + LoRA (screening JSON / officer memo) |
41
+ | Base | Qwen2.5-7B-Instruct 4-bit (MLX) |
42
+ | Method | MLX LoRA, rank 16, `mask_prompt` |
43
+ | Pick | v9-160 (ScreenBench, not min val loss) |
44
+ | Lake | Public SOP 50 10 / 50 55 / 50 57, 13 CFR 120–121, SBA notices — **retrieved**, not CPT |
45
+ | Mouth (optional) | v17-160 English; JSON still TRACE-bound |
46
+
47
+ ## Evaluation
48
+
49
+ Grader is code (`sba_screen/quality.py`). Not an LLM judge.
50
+
51
+ **ScreenBench holdout n=12**
52
+
53
+ | | All-pass | False-kill |
54
+ |---|---|---|
55
+ | This adapter, no TRACE bind | 6/12 | 0 |
56
+ | Complete system (lake + TRACE bind) | 12/12 | 0 |
57
+
58
+ Exam-v1 (n=76) with bind: all-pass 76/76, false-kill 0. Retrieve on occupancy / WC / clock sentinels: 13/13.
59
+
60
+ Do not quote 12/12 as a naked-7B number. The adapter copies the board. TRACE computes the board.
61
+
62
+ ## Quickstart
63
 
64
  ```python
65
  from mlx_lm import load, generate
66
+
67
  model, tokenizer = load(
68
  "mlx-community/Qwen2.5-7B-Instruct-4bit",
69
  adapter_path="AmPac/trace",
70
  )
71
  ```
72
 
73
+ Staff runner (opens 120.131 on investor apartments; no GPU required for TRACE-only):
74
 
75
  ```bash
76
+ python3 -m sba_screen --as-of 2026-01-15 --program 504 \
77
+ --sketch "504 to buy apartments they will rent out and not occupy"
78
  ```
79
 
80
+ Talk:
 
 
 
 
 
81
 
82
+ ```bash
83
+ python3 -m sba_screen --talk --as-of 2026-01-15 --program 504
84
+ ```
85
 
86
+ Studio mouth (`--mouth`) is optional English. TRACE still binds `hard_stop`.
87
 
88
+ ## Limitations
89
 
90
+ Does not determine eligibility. Does not decide credit. Does not speak for SBA. Trace is not an SBA product. SBA does not approve or endorse Trace. Do not give Trace to applicants. Outputs are drafts.
91
 
92
+ Not included in training: borrower NPI, tax returns, AmPac files, approve/decline tapes.
93
 
94
  ## License
95
 
96
+ `other`. Qwen2.5 weights are Tongyi Qianwen. The SOP/CFR lake is U.S. government work. Screening code is ours.