Trace card: README.md

README.md

@@ -10,6 +10,41 @@ tags:
   - lora
   - mlx
 pipeline_tag: text-generation
+model-index:
+- name: Trace
+  results:
+  - task:
+      type: text-generation
+      name: ScreenBench
+    dataset:
+      type: screenbench
+      name: ScreenBench holdout
+      split: holdout
+    metrics:
+    - type: hard_stop_all_pass
+      value: 1.0
+      name: Trace all-pass
+      verified: false
+    - type: false_kill_rate
+      value: 0.0
+      name: False kill
+      verified: false
+  - task:
+      type: text-generation
+      name: ScreenBench
+    dataset:
+      type: screenbench
+      name: ScreenBench exam-v1
+      split: exam
+    metrics:
+    - type: hard_stop_all_pass
+      value: 1.0
+      name: Trace all-pass
+      verified: false
+    - type: false_kill_rate
+      value: 0.0
+      name: False kill
+      verified: false
 ---
 
 # Trace
@@ -18,20 +53,77 @@ pipeline_tag: text-generation
 
 Trace is the 504/7(a) model for lenders.
 
-CDCs, CDFIs, and 7(a) shops use it the way they already use a model: paste a file, get an answer. The answer has to be the SOP or CFR paragraph that is in force on the loan-number date. If Trace cannot open that paragraph, it is not a screen.
+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.
 
-A person still owns the file. Trace does not approve. Trace does not deny. Trace is not an SBA product. Not for applicants.
+Trace does not approve. Trace does not deny. Trace is not an SBA product. Not for applicants. Outputs are drafts.
+
+![ScreenBench holdout](holdout.png)
+
+## What you are running
+
+Trace is not a naked 7B. It is three pieces that do not get averaged:
+
+![Tools, retrieve, Trace](columns.png)
+
+| Must open | File |
+|---|---|
+| **13 CFR 120.131** | Investor real estate, will not occupy |
+| **13 CFR 120.120** | Working capital on a 504 |
+| **SOP 50 10 8.1** | Loan number on 2026-10-15 |
+
+No cite → not a screen.
+
+## Quickstart
 
 ```bash
 python3 -m sba_screen --as-of 2026-01-15 --sketch "504 to buy apartments they will rent out and not occupy"
 ```
 
-- Investor real estate → **13 CFR 120.131**
-- Working capital on a 504 → **13 CFR 120.120**
-- Loan number on 2026-10-15 → **SOP 50 10 8.1**
+Base weights: [`mlx-community/Qwen2.5-7B-Instruct-4bit`](https://huggingface.co/mlx-community/Qwen2.5-7B-Instruct-4bit). This repo is the LoRA.
 
-Holdout n=12: Trace **12/12**. Weights alone 6/12. Same grader. `verified: false`.
+## ScreenBench
 
----
+Same code grader on every seat. `verified: false`.
+
+**Holdout n=12** (the published gate)
+
+| Seat | All-pass | False kill |
+|---|---|---|
+| Qwen 7B base | 0.00 | 0 |
+| Claude | 0.25 | 0 |
+| Weights alone | 0.50 | 0 |
+| **Trace** | **1.00** | **0** |
+
+**Exam-v1 n=76** (not the Hub gate)
+
+![Exam-v1](exam.png)
+
+| Seat | All-pass | False kill |
+|---|---|---|
+| Qwen 7B base | 0.00 | 0.12 |
+| Qwen + dated RAG | 0.16 | 0.04 |
+| Qwen + RAG + tools | 0.20 | 0.03 |
+| Weights alone | 0.93 | 0 |
+| **Trace** | **1.00** | **0** |
+
+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.
+
+## Intended use
+
+Lender and CDC staff screening 504 and 7(a) origination files. Paste a sketch. Open the paragraph. A designated officer certifies eligibility and credit.
+
+## Out of scope
+
+Consumer “am I eligible.” Credit decisions. Disaster, microloan, PPP, or Express as the same weights. An SBA system of record.
+
+## Model
+
+| | |
+|---|---|
+| Type | LoRA on Qwen2.5-7B-Instruct |
+| Rank / scale | 16 / 20 |
+| Pick | checkpoint 160 of 200 |
+| Lake | 23 official SOP/CFR/notice sources, retrieved, not CPT |
+| `adapters.safetensors` | sha256 `a7ef52fe9c4881e3904343dca7f75ee9ba01132d35de26bac424d3470a7f671b` |
 
-LoRA on [Qwen2.5-7B-Instruct-4bit](https://huggingface.co/mlx-community/Qwen2.5-7B-Instruct-4bit). Official SOP/CFR is retrieved, not trained in. License `other` (Qwen + U.S. government works + runner).
+License `other`: Qwen (Tongyi Qianwen) + U.S. government works as retrieved context + runner.