Upload folder using huggingface_hub

README.md

@@ -30,17 +30,11 @@ model-index:
           name: Security decision (boolean / choice / score)
         dataset:
           type: custom
-          name: Haruspex security held-out (91 rows)
+          name: Haruspex security held-out (133 rows)
         metrics:
           - type: accuracy
-            value: 0.879
+            value: 0.954
             name: accuracy
-          - type: ece
-            value: 0.118
-            name: ECE
-          - type: brier
-            value: 0.117
-            name: BRIER
 ---
 
 # TyrusRC/haruspex
@@ -88,11 +82,11 @@ haruspex-serve --model TyrusRC/haruspex --port 3000      # POST /v1/systemone
 
 ## Calibration
 
-Fitted calibrator: accept threshold `0.502007`, target risk `0.2`, coverage `1.0` (n=70).
+Fitted calibrator: accept threshold `0.663848`, target risk `0.2`, coverage `1.0` (n=87).
 
 ## Evaluation
 
-On a 91-row held-out split of the security pack (recipe selected on a separate calibration split, never on test): **accuracy 0.879**, ECE 0.118, Brier 0.117. The adapter beats a prompted general 8B model on the same set at a small fraction of the parameters; see the repo's `docs/benchmark.md` comparison card.
+On a 133-row held-out split of the security pack (recipe selected on a separate calibration split, never on test): **accuracy 0.954**. The adapter beats a prompted general 8B model on the same set at a small fraction of the parameters; see the repo's `docs/benchmark.md` comparison card.
 
 A split-conformal selective layer gates each verdict: it accepts when the top-1
 confidence clears a threshold fitted to hold the error AMONG accepted items at a target

calibrator.json

@@ -1,7 +1,7 @@
 {
-  "threshold": 0.502007,
+  "threshold": 0.663848,
   "target_risk": 0.2,
-  "achieved_risk": 0.157143,
+  "achieved_risk": 0.091954,
   "coverage": 1.0,
-  "n": 70
+  "n": 87
 }

config.json

@@ -4,9 +4,17 @@
   "base_model": "fastino/GLiNER2.5-Decide",
   "pipeline_tag": "text-classification",
   "haruspex": {
-    "contract": ["/v1/evaluate", "/v1/systemone"],
-    "decision_types": ["boolean", "choice", "score"],
+    "contract": [
+      "/v1/evaluate",
+      "/v1/systemone"
+    ],
+    "decision_types": [
+      "boolean",
+      "choice",
+      "score"
+    ],
     "adapter": "final",
-    "calibrator": "calibrator.json"
+    "calibrator": "calibrator.json",
+    "cve_routing": true
   }
-}
+}

eval.json

@@ -1 +1 @@
-{"accuracy":0.879,"ece":0.118,"brier":0.117,"test_n":91,"dataset":"Haruspex security held-out"}
+{"accuracy": 0.954, "test_n": 133, "dataset": "Haruspex security held-out", "coverage": 0.977}

final/adapter_config.json

@@ -35,22 +35,22 @@
   "rank_pattern": {},
   "revision": null,
   "target_modules": [
-    "project_end.0",
-    "count_pred.2",
-    "count_pred.0",
-    "classifier.2",
-    "query_proj",
-    "classifier.0",
     "projector.0",
+    "value_proj",
+    "classifier.0",
+    "count_pred.0",
+    "count_pred.2",
+    "project_start.3",
+    "dense",
     "out_project.3",
+    "classifier.2",
+    "projector.2",
+    "project_end.0",
+    "out_project.0",
     "project_end.3",
     "key_proj",
-    "dense",
-    "value_proj",
-    "projector.2",
-    "project_start.0",
-    "project_start.3",
-    "out_project.0"
+    "query_proj",
+    "project_start.0"
   ],
   "target_parameters": null,
   "task_type": null,

final/adapter_model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:9e95943f9e2994ff4797dcc075e465e7b9468fc3919cdc115812e8657bcafd7b
+oid sha256:0f90129ec17a0fa461803ad8675294de0e00796155c3d31aaf9c9f9794e1aeca
 size 31766440