Upload GNN turn-level model artifacts

README.md

@@ -13,19 +13,19 @@ model-index:
         metrics:
           - name: F1
             type: f1
-            value: 0.2062
+            value: 0.8586
           - name: PR-AUC
             type: pr_auc
-            value: 0.2205
+            value: 0.9720
           - name: ROC-AUC
             type: roc_auc
-            value: 0.8497
+            value: 0.9772
           - name: Precision
             type: precision
-            value: 0.1437
+            value: 0.8589
           - name: Recall
             type: recall
-            value: 0.3744
+            value: 0.9158
 ---
 # GNN Jailbreak Prediction Model (phi4:14b)
 
@@ -35,12 +35,12 @@ Homogeneous GNN classifier for unsafe/jailbreak likelihood in multi-turn convers
 
 | Metric         | Value  |
 |----------------|--------|
-| F1             | 0.2062 |
-| PR-AUC         | 0.2205 |
-| ROC-AUC        | 0.8497 |
-| Precision      | 0.1437 |
-| Recall         | 0.3744 |
-| Best Threshold | 0.780 |
+| F1             | 0.8586 |
+| PR-AUC         | 0.9720 |
+| ROC-AUC        | 0.9772 |
+| Precision      | 0.8589 |
+| Recall         | 0.9158 |
+| Best Threshold | 0.390 |
 
 ## Training Details
 
@@ -55,4 +55,4 @@ Homogeneous GNN classifier for unsafe/jailbreak likelihood in multi-turn convers
 
 ## Dataset Size (training samples)
 
-Prepared turn-level samples: 2585
+Prepared turn-level samples: 395

gnn_homo_payload.pt

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:10bcdd0597c5bdbaaf39e8d99347ce740a5f789e9b2e3d32408dd1ad69c3987d
-size 1045189
+oid sha256:edb1c160195c31c644ebb39468cbeed6ecfb41393cebc20f32c681f6161be870
+size 971461

metadata.json

@@ -1,16 +1,16 @@
 {
   "csv": "/home/digayona/multi_turn_jailbreak_RL/GNN/turns_table_llama3_8b_harmbench.csv",
   "target_model": "phi4:14b",
-  "threshold": 0.78,
+  "threshold": 0.39,
   "sentence_model_name": "sentence-transformers/all-MiniLM-L6-v2",
-  "n_rows": 2585,
+  "n_rows": 395,
   "n_models": 1,
   "split_col": "goal",
   "seed": 42,
   "turn_norm_mode": "dataset_max",
   "turn_norm_denom": 22.0,
   "session_len_norm_mode": "dataset_max",
-  "session_len_norm_denom": 22.0,
+  "session_len_norm_denom": 20.0,
   "model_kwargs": {
     "hidden_channels": 128,
     "num_layers": 2,
@@ -20,10 +20,10 @@
   },
   "use_turn_bucket_features": false,
   "test_metrics": {
-    "roc_auc": 0.8497476977719763,
-    "pr_auc": 0.22046366741105397,
-    "f1": 0.2062108262108262,
-    "precision": 0.1437456148001771,
-    "recall": 0.3743833943833944
+    "roc_auc": 0.9772275091195899,
+    "pr_auc": 0.9720258299076259,
+    "f1": 0.8585849597195537,
+    "precision": 0.8589285714285715,
+    "recall": 0.9158333333333333
   }
 }