Upload NLIScorer

pipeline.py

@@ -331,31 +331,31 @@ TASK_CLASSES = {
 }
 
 TASK_THRESHOLDS = {
-    "Detection/Hallucination/Factual Consistency": 0.5895,
-    "Detection/Prompt Injection": 0.4147,
-    "Detection/Source Code": 0.4001,
-    "Detection/Toxicity/Disability": 0.5547,
-    "Detection/Toxicity/Gender": 0.4007,
-    "Detection/Toxicity/Identity Hate": 0.5502,
-    "Detection/Toxicity/Insult": 0.4913,
-    "Detection/Toxicity/Obscene": 0.448,
-    "Detection/Toxicity/Race": 0.5983,
-    "Detection/Toxicity/Religion": 0.4594,
-    "Detection/Toxicity/Toxic": 0.5034,
-    "Detection/Toxicity/Violence": 0.4031,
-    "Quality/Context/Document Relevance": 0.5809,
-    "Quality/Context/Document Utilization": 0.4005,
-    "Quality/Context/Sentence Relevance": 0.6003,
-    "Quality/Context/Sentence Utilization": 0.5417,
-    "Quality/Response/Adherence": 0.59,
-    "Quality/Response/Attribution": 0.5304,
-    "Quality/Response/Coherence": 0.6891,
-    "Quality/Response/Complexity": 0.7235,
-    "Quality/Response/Correctness": 0.6535,
-    "Quality/Response/Helpfulness": 0.4445,
-    "Quality/Response/Instruction Following": 0.5323,
-    "Quality/Response/Relevance": 0.4011,
-    "Quality/Response/Verbosity": 0.4243,
+    "Detection/Hallucination/Factual Consistency": 0.5,
+    "Detection/Prompt Injection": 0.5001,
+    "Detection/Source Code": 0.5039,
+    "Detection/Toxicity/Disability": 0.5111,
+    "Detection/Toxicity/Gender": 0.5003,
+    "Detection/Toxicity/Identity Hate": 0.5035,
+    "Detection/Toxicity/Insult": 0.5187,
+    "Detection/Toxicity/Obscene": 0.5034,
+    "Detection/Toxicity/Race": 0.5081,
+    "Detection/Toxicity/Religion": 0.5058,
+    "Detection/Toxicity/Toxic": 0.5005,
+    "Detection/Toxicity/Violence": 0.5001,
+    "Quality/Context/Document Relevance": 0.5016,
+    "Quality/Context/Document Utilization": 0.5014,
+    "Quality/Context/Sentence Relevance": 0.5002,
+    "Quality/Context/Sentence Utilization": 0.5039,
+    "Quality/Response/Adherence": 0.5107,
+    "Quality/Response/Attribution": 0.5053,
+    "Quality/Response/Coherence": 0.6103,
+    "Quality/Response/Complexity": 0.5603,
+    "Quality/Response/Correctness": 0.501,
+    "Quality/Response/Helpfulness": 0.5018,
+    "Quality/Response/Instruction Following": 0.5001,
+    "Quality/Response/Relevance": 0.5012,
+    "Quality/Response/Verbosity": 0.5408,
 }
 
 
@@ -381,12 +381,8 @@ class NLIScorer(Pipeline):
         return outputs
 
     def postprocess(self, model_outputs, task_type=None, threshold=None):
-        if threshold is not None:
-            threshold = threshold
-        elif task_type is not None:
-            threshold = TASK_THRESHOLDS[task_type]
-        else:
-            threshold = 0.5
+        if threshold is None:
+            threshold = TASK_THRESHOLDS.get(task_type, 0.5)
         pos_scores = model_outputs["logits"].softmax(-1)[0][1]
         best_class = int(pos_scores > threshold)
         if best_class == 1: