Upload README.md with huggingface_hub
Browse files
README.md
CHANGED
|
@@ -57,3 +57,12 @@ logits, _, _ = lens.apply(model, "Fact: the currency of the country shaped like
|
|
| 57 |
|
| 58 |
Findings, caveats, and every analysis script:
|
| 59 |
[github.com/solarkyle/jspace](https://github.com/solarkyle/jspace)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 57 |
|
| 58 |
Findings, caveats, and every analysis script:
|
| 59 |
[github.com/solarkyle/jspace](https://github.com/solarkyle/jspace)
|
| 60 |
+
|
| 61 |
+
## Router weights (`router/`)
|
| 62 |
+
|
| 63 |
+
`workspace_router_e4b.json` - the trained hallucination-risk classifier
|
| 64 |
+
(logistic regression, 10 workspace features / 14 combined) from the repo's
|
| 65 |
+
`analyze_router.py`. Trained on E4B, transfers zero-shot to the other Gemmas
|
| 66 |
+
when features are z-scored per model (AUC 0.74-0.78). Predicts P(wrong);
|
| 67 |
+
escalate when risk exceeds your budget threshold. Feature definitions in
|
| 68 |
+
[analyze_router.py](https://github.com/solarkyle/jspace/blob/master/analyze_router.py).
|