Hugging Face's logo

AmPac
/
trace

Text Generation
MLX
English
sba
504
7a
lora
screening
Model card Files
xet
 Community
mzwin9870 commited on
Commit
2519bb9
·
verified ·
1 Parent(s): 9550e4e

Trace card: README.md — clock, sources, training graph

Browse files
Files changed (1) hide show
  1. README.md +56 -60
README.md CHANGED
@@ -29,63 +29,75 @@ model-index:
29
  value: 0.0
30
  name: False kill
31
  verified: false
32
- - task:
33
- type: text-generation
34
- name: ScreenBench
35
- dataset:
36
- type: screenbench
37
- name: ScreenBench exam-v1
38
- split: exam
39
- metrics:
40
- - type: hard_stop_all_pass
41
- value: 1.0
42
- name: Trace all-pass
43
- verified: false
44
- - type: false_kill_rate
45
- value: 0.0
46
- name: False kill
47
- verified: false
48
  ---
49
 
50
  # Trace
51
 
52
  **Occupancy opens 13 CFR 120.131.**
53
 
54
- Trace is the 504/7(a) model for lenders.
 
 
55
 
56
- CDCs, CDFIs, and 7(a) shops already paste files into a model. Trace is that model for this job. The answer has to be the SOP or CFR paragraph in force on the loan-number date. A person still owns the file.
57
 
58
- Trace does not approve. Trace does not deny. Trace is not an SBA product. Not for applicants. Outputs are drafts.
59
 
60
- ![ScreenBench holdout](holdout.png)
61
 
62
- ## What you are running
 
 
 
 
63
 
64
- Trace is not a naked 7B. It is three pieces that do not get averaged:
65
 
66
- ![Tools, retrieve, Trace](columns.png)
67
 
68
- | Must open | File |
69
- |---|---|
70
- | **13 CFR 120.131** | Investor real estate, will not occupy |
71
- | **13 CFR 120.120** | Working capital on a 504 |
72
- | **SOP 50 10 8.1** | Loan number on 2026-10-15 |
 
73
 
74
  No cite → not a screen.
75
 
76
- ## Quickstart
77
 
78
- ```bash
79
- python3 -m sba_screen --as-of 2026-01-15 --sketch "504 to buy apartments they will rent out and not occupy"
80
- ```
 
 
 
 
 
 
 
 
 
 
 
 
 
81
 
82
- Base weights: [`mlx-community/Qwen2.5-7B-Instruct-4bit`](https://huggingface.co/mlx-community/Qwen2.5-7B-Instruct-4bit). This repo is the LoRA.
 
 
 
 
 
 
83
 
84
  ## ScreenBench
85
 
86
- Same code grader on every seat. `verified: false`.
87
 
88
- **Holdout n=12** (the published gate)
 
 
89
 
90
  | Seat | All-pass | False kill |
91
  |---|---|---|
@@ -96,34 +108,18 @@ Same code grader on every seat. `verified: false`.
96
 
97
  **Exam-v1 n=76** (not the Hub gate)
98
 
99
- ![Exam-v1](exam.png)
100
 
101
- | Seat | All-pass | False kill |
102
- |---|---|---|
103
- | Qwen 7B base | 0.00 | 0.12 |
104
- | Qwen + dated RAG | 0.16 | 0.04 |
105
- | Qwen + RAG + tools | 0.20 | 0.03 |
106
- | Weights alone | 0.93 | 0 |
107
- | **Trace** | **1.00** | **0** |
108
-
109
- The 12/12 number is Trace the product. The 6/12 number is the adapter file with the runner off. Both stay on this card.
110
-
111
- ## Intended use
112
-
113
- Lender and CDC staff screening 504 and 7(a) origination files. Paste a sketch. Open the paragraph. A designated officer certifies eligibility and credit.
114
 
115
- ## Out of scope
116
 
117
- Consumer “am I eligible.” Credit decisions. Disaster, microloan, PPP, or Express as the same weights. An SBA system of record.
 
 
118
 
119
- ## Model
120
 
121
- | | |
122
- |---|---|
123
- | Type | LoRA on Qwen2.5-7B-Instruct |
124
- | Rank / scale | 16 / 20 |
125
- | Pick | checkpoint 160 of 200 |
126
- | Lake | 23 official SOP/CFR/notice sources, retrieved, not CPT |
127
- | `adapters.safetensors` | sha256 `a7ef52fe9c4881e3904343dca7f75ee9ba01132d35de26bac424d3470a7f671b` |
128
 
129
- License `other`: Qwen (Tongyi Qianwen) + U.S. government works as retrieved context + runner.
 
29
  value: 0.0
30
  name: False kill
31
  verified: false
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
32
  ---
33
 
34
  # Trace
35
 
36
  **Occupancy opens 13 CFR 120.131.**
37
 
38
+ Trace is the 504/7(a) model for lenders. CDCs, CDFIs, and 7(a) shops paste a file and get the paragraph in force. A person still owns the file. Trace does not approve. Trace does not deny. Trace is not an SBA product. Not for applicants.
39
+
40
+ ![ScreenBench holdout](holdout.png)
41
 
42
+ ## The clock (this is not a leaked loan)
43
 
44
+ SBA numbers an application on a calendar day. That day — `as_of` — picks which SOP governs the file. It is **not** a borrower ID and not an AmPac loan.
45
 
46
+ ![SOP clock](clock.png)
47
 
48
+ | `as_of` | SOP in force | Official text |
49
+ |---|---|---|
50
+ | through 2025-05-31 | SOP 50 10 7.1 | [docx](https://www.sba.gov/sites/default/files/2023-10/SOP%2050%2010%207.1%20effective%2011.15.23.docx) |
51
+ | 2025-06-01 to 2026-09-30 | SOP 50 10 8 | [docx](https://www.sba.gov/sites/default/files/2025-05/SOP%2050%2010%208%20Technical%20Updates%20effective%206.1.2025.docx) |
52
+ | on or after 2026-10-01 | SOP 50 10 8.1 | [docx](https://www.sba.gov/sites/default/files/2026-08/SOP%2050%2010%208.1%20effective%2010.1.2026_0.docx) · [notice issuing 8.1](https://www.sba.gov/sites/default/files/2026-08/SBA%20Information%20Notice%205000-880695%20-%20Issuance%20of%20SOP%2050%2010%208.1.pdf) |
53
 
54
+ So **as_of 2026-10-15 → SOP 50 10 8.1**. That is a test date after the 8.1 cutover. Clock is code (`clock.py`). The SOP body is retrieved, not memorized.
55
 
56
+ ## What must open
57
 
58
+ | Sketch | Must open | Official |
59
+ |---|---|---|
60
+ | Investor real estate, will not occupy | 13 CFR 120.131 | [eCFR](https://www.ecfr.gov/current/title-13/section-120.131) |
61
+ | Working capital / payroll on a 504 | 13 CFR 120.120 | [eCFR](https://www.ecfr.gov/current/title-13/section-120.120) |
62
+ | Size (NAICS + receipts) | 13 CFR 121.201 | [eCFR](https://www.ecfr.gov/current/title-13/section-121.201) · [size table](https://data.sba.gov/sites/default/files/distribution/SBA-GCBD-2014-08-001/sba-table-of-size-standards_effective-march-17-2023_v0.xlsx) |
63
+ | as_of 2026-10-15 | SOP 50 10 8.1 | [SBA SOP 50 10](https://www.sba.gov/document/sop-50-10-lender-development-company-loan-programs) |
64
 
65
  No cite → not a screen.
66
 
67
+ ## Official sources Trace retrieves
68
 
69
+ Public U.S. government works. Dated snapshots. **Retrieved context, not a continued-pretraining dump.**
70
+
71
+ | Source | In force | Link |
72
+ |---|---|---|
73
+ | 13 CFR Part 120 | current eCFR | [ecfr.gov](https://www.ecfr.gov/current/title-13/chapter-I/part-120) |
74
+ | 13 CFR Part 121 (size) | current eCFR | [ecfr.gov](https://www.ecfr.gov/current/title-13/chapter-I/part-121) |
75
+ | SOP 50 10 8 | 2025-06-01 | [SBA](https://www.sba.gov/sites/default/files/2025-05/SOP%2050%2010%208%20Technical%20Updates%20effective%206.1.2025.docx) |
76
+ | SOP 50 10 8.1 | 2026-10-01 | [SBA](https://www.sba.gov/sites/default/files/2026-08/SOP%2050%2010%208.1%20effective%2010.1.2026_0.docx) |
77
+ | SOP 50 10 7.1 | historical, through 2025-05-31 | [SBA](https://www.sba.gov/sites/default/files/2023-10/SOP%2050%2010%207.1%20effective%2011.15.23.docx) |
78
+ | SOP 50 55 (504 servicing) | 2013-10-01 | [SBA PDF](https://www.sba.gov/sites/default/files/2022-07/SOP_50_55_9-4-13_1.pdf) |
79
+ | SOP 50 57 (7(a) servicing) | 2025-11-01 | [SBA PDF](https://www.sba.gov/sites/default/files/2025-09/FINAL%20SOP%205057%204%20version.pdf) |
80
+ | PN 5000-876441 citizenship | 2026-03-01 | [SBA PDF](https://www.sba.gov/sites/default/files/2026-02/Policy%20Notice%205000-876441%20-%20Update%20to%20Citizenship%20and%20Residency%20Requirements%20and%20Recission%20of%20PN%2050000-872050.pdf) |
81
+ | IN 5000-868665 (issues SOP 8) | 2025-05-29 | [SBA PDF](https://www.sba.gov/sites/default/files/2025-05/SBA%20Information%20Notice%205000-868665_0.pdf) |
82
+ | IN 5000-880695 (issues SOP 8.1) | 2026-08-14 | [SBA PDF](https://www.sba.gov/sites/default/files/2026-08/SBA%20Information%20Notice%205000-880695%20-%20Issuance%20of%20SOP%2050%2010%208.1.pdf) |
83
+ | Size standards table | 2023-03-17 | [data.sba.gov xlsx](https://data.sba.gov/sites/default/files/distribution/SBA-GCBD-2014-08-001/sba-table-of-size-standards_effective-march-17-2023_v0.xlsx) |
84
+ | Franchise directory | 2026-08-20 | [SBA xlsx](https://www.sba.gov/sites/default/files/2026-08/Franchise%20Directory%20Aug%2020-2026-REM01.xlsx) |
85
 
86
+ Lake on disk: 23 official sources, 4,775 chunks. Size and franchise are **lookups**, not 978 RAG chunks.
87
+
88
+ ## What we trained
89
+
90
+ ![What went into Trace](train.png)
91
+
92
+ Weights learn **how to screen**. Facts that move (SOP revs, notices) stay in the dated lake. If it is only in the weights and the SOP changes, the model lies with confidence.
93
 
94
  ## ScreenBench
95
 
96
+ Same code grader. `verified: false`.
97
 
98
+ ![Tools / retrieve / Trace](columns.png)
99
+
100
+ **Holdout n=12**
101
 
102
  | Seat | All-pass | False kill |
103
  |---|---|---|
 
108
 
109
  **Exam-v1 n=76** (not the Hub gate)
110
 
111
+ ![Exam](exam.png)
112
 
113
+ 12/12 is Trace the product (weights + lake + runner). 6/12 is the adapter with the runner off.
 
 
 
 
 
 
 
 
 
 
 
 
114
 
115
+ ## Quickstart
116
 
117
+ ```bash
118
+ python3 -m sba_screen --as-of 2026-01-15 --sketch "504 to buy apartments they will rent out and not occupy"
119
+ ```
120
 
121
+ That must open [13 CFR 120.131](https://www.ecfr.gov/current/title-13/section-120.131).
122
 
123
+ ## License
 
 
 
 
 
 
124
 
125
+ `other`. Qwen2.5 (Tongyi Qianwen) + U.S. government SOP/CFR as retrieved context + AmPac runner. Official text was not a continued-pretraining target. Live borrower files were not training data.