mzwin9870 commited on
Commit
4ce759f
·
verified ·
1 Parent(s): 3afb7ec

alpha card: occupancy first, drop 12/12 scoreboard

Browse files
Files changed (1) hide show
  1. README.md +31 -78
README.md CHANGED
@@ -10,73 +10,44 @@ tags:
10
  - lora
11
  - mlx
12
  pipeline_tag: text-generation
13
- model-index:
14
- - name: Trace
15
- results:
16
- - task:
17
- type: text-generation
18
- name: ScreenBench
19
- dataset:
20
- type: screenbench
21
- name: ScreenBench holdout
22
- split: holdout
23
- metrics:
24
- - type: hard_stop_all_pass
25
- value: 1.0
26
- name: Trace all-pass
27
- verified: false
28
- - type: false_kill_rate
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 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.
39
 
40
- Not an SBA product. Does not speak for SBA. Does not determine eligibility. Does not decide credit. Not for applicants.
41
 
42
- ## ScreenBench
43
 
44
- Same 12 files. Same grader. `verified: false`. Frontier seats are naked JSON. Trace is the product seat.
45
 
46
- ![Frontier vs Trace](compare.png)
 
 
47
 
48
- ![Who opens which kill](families.png)
49
 
50
- | Seat | All-pass | False kill | SOP clock |
51
- |---|---|---|---|
52
- | Qwen 2.5 7B | 0.00 | 0 | 1.00 |
53
- | gpt-oss 20B | 0.00 | 0.08 | 1.00 |
54
- | Gemma 4 31B | 0.00 | 0 | 1.00 |
55
- | Claude | 0.08 | 0.08 | 0.75 |
56
- | Codex | 0.25 | 0.08 | 0.50 |
57
- | Grok | 0.08 | 0.08 | 0.83 |
58
- | **Trace** | **1.00** | **0** | **1.00** |
59
 
60
- Claude 1/12 (owner-user-pass). Grok 1/12 (occupancy). Codex 3/12 (occupancy, affiliation, owner-user-pass). Trace 12/12 product seat. Naked adapter 0.50. v0 bar 0.85. No seat except Trace clears. `verified: false`. Identification, not a partnership.
61
 
62
  ## SOP in force
63
 
64
- The loan-number date picks the SOP. SOP 50 10 8.1 does not leak onto a January 2026 file.
65
 
66
  ![SOP clock](clock.png)
67
 
68
- ![SOP 8.1 leak](clock-leak.png)
69
-
70
- Claude put SOP 50 10 8.1 on 3 of 11 files numbered before 2026-10-01. Codex 6 of 11. Grok 2 of 11. Trace 0 of 11. Clock is code.
71
-
72
  | Loan-number date | SOP | Official |
73
  |---|---|---|
74
  | through 2025-05-31 | SOP 50 10 7.1 | [SBA SOP 50 10](https://www.sba.gov/document/sop-50-10-lender-development-company-loan-programs) · [docx](https://legacy.sba.gov/sites/default/files/2023-10/SOP%2050%2010%207.1%20effective%2011.15.23.docx) |
75
  | 2025-06-01 to 2026-09-30 | SOP 50 10 8 | [SBA SOP 50 10](https://www.sba.gov/document/sop-50-10-lender-development-company-loan-programs) · [docx](https://legacy.sba.gov/sites/default/files/2025-05/SOP%2050%2010%208%20Technical%20Updates%20effective%206.1.2025.docx) |
76
  | on or after 2026-10-01 | SOP 50 10 8.1 | [SBA SOP 50 10](https://www.sba.gov/document/sop-50-10-lender-development-company-loan-programs) · [docx](https://legacy.sba.gov/sites/default/files/2026-08/SOP%2050%2010%208.1%20effective%2010.1.2026_0.docx) · [notice](https://www.sba.gov/document/information-notice-5000-880695-issuance-sop-50-10-81) |
77
 
78
- Clock is code. The SOP body is retrieved, not memorized.
79
-
80
  ## What must open
81
 
82
  | Sketch | Must open | Official |
@@ -86,11 +57,11 @@ Clock is code. The SOP body is retrieved, not memorized.
86
  | Size (NAICS + receipts) | 13 CFR 121.201 | [eCFR](https://www.ecfr.gov/current/title-13/section-121.201) · [size table](https://www.sba.gov/document/support-table-size-standards) |
87
  | Loan numbered on or after 2026-10-01 | SOP 50 10 8.1 | [SBA SOP 50 10](https://www.sba.gov/document/sop-50-10-lender-development-company-loan-programs) |
88
 
89
- No cite → not a screen.
90
 
91
- ## Official sources Trace retrieves
92
 
93
- Public U.S. government works. Dated snapshots. **Retrieved context, not a continued-pretraining dump.** Direct file links use `legacy.sba.gov` because SBA's `www` file paths 404.
94
 
95
  | Source | In force | Official |
96
  |---|---|---|
@@ -102,46 +73,28 @@ Public U.S. government works. Dated snapshots. **Retrieved context, not a contin
102
  | SOP 50 55 (504 servicing) | 2013-10-01 | [page](https://www.sba.gov/document/sop-50-55-504-loan-servicing-liquidation) · [pdf](https://legacy.sba.gov/sites/default/files/2022-07/SOP_50_55_9-4-13_1.pdf) |
103
  | SOP 50 57 (7(a) servicing) | 2025-11-01 | [page](https://www.sba.gov/document/sop-50-57-7a-loan-servicing-liquidation) · [pdf](https://legacy.sba.gov/sites/default/files/2025-09/FINAL%20SOP%205057%204%20version.pdf) |
104
  | PN 5000-876441 citizenship | 2026-03-01 | [pdf](https://legacy.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) |
105
- | IN 5000-868665 (issues SOP 8) | 2025-05-29 | [pdf](https://legacy.sba.gov/sites/default/files/2025-05/SBA%20Information%20Notice%205000-868665_0.pdf) |
106
- | IN 5000-880695 (issues SOP 8.1) | 2026-08-14 | [page](https://www.sba.gov/document/information-notice-5000-880695-issuance-sop-50-10-81) · [pdf](https://legacy.sba.gov/sites/default/files/2026-08/SBA%20Information%20Notice%205000-880695%20-%20Issuance%20of%20SOP%2050%2010%208.1.pdf) |
107
- | Size standards table | 2023-03-17 | [page](https://www.sba.gov/document/support-table-size-standards) · [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) |
108
- | Franchise directory | 2026-08-20 | [page](https://www.sba.gov/document/support-sba-franchise-directory) · [xlsx](https://legacy.sba.gov/sites/default/files/2026-08/Franchise%20Directory%20Aug%2020-2026-REM01.xlsx) |
109
-
110
- Lake on disk: 23 official sources, 4,775 chunks. Size and franchise are **lookups**, not 978 RAG chunks.
111
-
112
- ## Three columns. Never averaged.
113
 
114
- ![Tools / retrieve / Trace](columns.png)
115
 
116
- 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.
117
 
118
- ## exam-v1 (not the Hub gate)
119
-
120
- Same 76 files. Same grader. Frontier seats are naked JSON. Trace is the product seat. Holdout n=12 is the Hub number.
121
-
122
- ![exam-v1 all-pass](exam-compare.png)
123
-
124
- ![exam-v1 families](exam-families.png)
125
-
126
- | Seat | All-pass | False kill | SOP clock |
127
- |---|---|---|---|
128
- | Qwen 2.5 7B | 0.00 (0/76) | 0.12 | 0.00 |
129
- | Qwen + dated RAG | 0.16 (12/76) | 0.04 | 0.70 |
130
- | Qwen + RAG + tools | 0.20 (15/76) | 0.03 | 0.71 |
131
- | Claude | 0.34 (26/76) | 0.05 | 0.66 |
132
- | Codex | 0.25 (19/76) | 0 | 0.89 |
133
- | Grok | 0.34 (26/76) | 0.09 | 1.00 |
134
- | **Trace** | **0.97 (74/76)** | **0** | **1.00** |
135
-
136
- Codex occupancy 8/8; pass family 0/20. Grok size 7/8; occupancy 0/8. Claude SOP 8.1 leak on 26 files. Trace affiliation 6/8; every other family 8/8 or 20/20. No frontier clears v0 0.85. `verified: false`.
137
-
138
- ## Quickstart
139
 
140
  ```bash
141
  python3 -m sba_screen --as-of 2026-01-15 --sketch "504 to buy apartments they will rent out and not occupy"
142
  ```
143
 
144
- That must open [13 CFR 120.131](https://www.ecfr.gov/current/title-13/section-120.131).
 
 
 
 
 
 
 
 
145
 
146
  ## License
147
 
 
10
  - lora
11
  - mlx
12
  pipeline_tag: text-generation
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
13
  ---
14
 
15
+ # Trace (alpha)
16
 
17
  **Occupancy opens 13 CFR 120.131.**
18
 
19
+ Trace is an **alpha** open-source 504/7(a) screening model for CDFIs, CDCs, and lenders. Paste a file sketch. It returns the SOP or CFR paragraph in force on the loan-number date, a draft hard-stop or bounce, and a board an officer can check.
20
 
21
+ A designated officer still certifies eligibility and credit. Trace does not approve a loan. Trace does not deny credit. Trace does not determine eligibility. Trace does not speak for SBA. Not an SBA product. Not for applicants. Outputs are drafts.
22
 
23
+ This is early. The unit is **weights + dated public lake + runner**. A naked adapter without the lake and TRACE is not Trace.
24
 
25
+ ## What you get
26
 
27
+ 1. **Open this** — the paragraph in force on `as_of` (the SBA loan-number date).
28
+ 2. **Draft hard-stop or bounce** — occupancy, ineligible type, working capital on a 504, packaging gaps.
29
+ 3. **Board** — clock through franchise, so a human can check the work.
30
 
31
+ If it cannot open a paragraph, it is not a screen.
32
 
33
+ ```bash
34
+ python3 -m sba_screen --as-of 2026-01-15 --sketch "504 to buy apartments they will rent out and not occupy"
35
+ ```
 
 
 
 
 
 
36
 
37
+ That must open [13 CFR 120.131](https://www.ecfr.gov/current/title-13/section-120.131). Payroll on a 504 must open [13 CFR 120.120](https://www.ecfr.gov/current/title-13/section-120.120).
38
 
39
  ## SOP in force
40
 
41
+ The loan-number date picks the SOP. SOP 50 10 8.1 does not leak onto a January 2026 file. Clock is code. The SOP body is retrieved, not memorized.
42
 
43
  ![SOP clock](clock.png)
44
 
 
 
 
 
45
  | Loan-number date | SOP | Official |
46
  |---|---|---|
47
  | through 2025-05-31 | SOP 50 10 7.1 | [SBA SOP 50 10](https://www.sba.gov/document/sop-50-10-lender-development-company-loan-programs) · [docx](https://legacy.sba.gov/sites/default/files/2023-10/SOP%2050%2010%207.1%20effective%2011.15.23.docx) |
48
  | 2025-06-01 to 2026-09-30 | SOP 50 10 8 | [SBA SOP 50 10](https://www.sba.gov/document/sop-50-10-lender-development-company-loan-programs) · [docx](https://legacy.sba.gov/sites/default/files/2025-05/SOP%2050%2010%208%20Technical%20Updates%20effective%206.1.2025.docx) |
49
  | on or after 2026-10-01 | SOP 50 10 8.1 | [SBA SOP 50 10](https://www.sba.gov/document/sop-50-10-lender-development-company-loan-programs) · [docx](https://legacy.sba.gov/sites/default/files/2026-08/SOP%2050%2010%208.1%20effective%2010.1.2026_0.docx) · [notice](https://www.sba.gov/document/information-notice-5000-880695-issuance-sop-50-10-81) |
50
 
 
 
51
  ## What must open
52
 
53
  | Sketch | Must open | Official |
 
57
  | Size (NAICS + receipts) | 13 CFR 121.201 | [eCFR](https://www.ecfr.gov/current/title-13/section-121.201) · [size table](https://www.sba.gov/document/support-table-size-standards) |
58
  | Loan numbered on or after 2026-10-01 | SOP 50 10 8.1 | [SBA SOP 50 10](https://www.sba.gov/document/sop-50-10-lender-development-company-loan-programs) |
59
 
60
+ ## Dated lake
61
 
62
+ Public U.S. government works. Snapshots on disk. **Retrieved context, not a continued-pretraining dump.** Direct file links use `legacy.sba.gov` because SBA `www` file paths 404.
63
 
64
+ 23 official sources. Size and franchise are **lookups**, not RAG.
65
 
66
  | Source | In force | Official |
67
  |---|---|---|
 
73
  | SOP 50 55 (504 servicing) | 2013-10-01 | [page](https://www.sba.gov/document/sop-50-55-504-loan-servicing-liquidation) · [pdf](https://legacy.sba.gov/sites/default/files/2022-07/SOP_50_55_9-4-13_1.pdf) |
74
  | SOP 50 57 (7(a) servicing) | 2025-11-01 | [page](https://www.sba.gov/document/sop-50-57-7a-loan-servicing-liquidation) · [pdf](https://legacy.sba.gov/sites/default/files/2025-09/FINAL%20SOP%205057%204%20version.pdf) |
75
  | PN 5000-876441 citizenship | 2026-03-01 | [pdf](https://legacy.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) |
76
+ | Size standards table | 2023-03-17 | [page](https://www.sba.gov/document/support-table-size-standards) |
77
+ | Franchise directory | 2026-08-20 | [page](https://www.sba.gov/document/support-sba-franchise-directory) |
 
 
 
 
 
 
78
 
79
+ Facts that move (SOP revs, notices) stay in the lake. If a fact lives only in the weights and the SOP changes, the model lies with confidence.
80
 
81
+ ## How to run
82
 
83
+ The paragraph has to open **without** treating the adapter as a chatbot.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
84
 
85
  ```bash
86
  python3 -m sba_screen --as-of 2026-01-15 --sketch "504 to buy apartments they will rent out and not occupy"
87
  ```
88
 
89
+ MLX LoRA on Qwen2.5-7B-Instruct-4bit is optional and runs on Apple Silicon. Base: [mlx-community/Qwen2.5-7B-Instruct-4bit](https://huggingface.co/mlx-community/Qwen2.5-7B-Instruct-4bit).
90
+
91
+ ## Alpha limits
92
+
93
+ - Does not determine eligibility.
94
+ - Does not decide credit.
95
+ - Does not speak for SBA. SBA does not approve or endorse Trace.
96
+ - Do not give Trace to applicants. Do not show the draft to a borrower.
97
+ - ScreenBench is an internal exam. It is not the product. `verified: false`.
98
 
99
  ## License
100