Merge branch 'shuguang/main' of https://github.com/katanemo/archgw into shuguang/main

model_server/src/core/function_calling.py

@@ -521,6 +521,7 @@ class ArchFunctionHandler(ArchBaseHandler):
         if has_tool_call is False:
             # [TODO] - Review: remove the following code
             print("No tool call found, start parameter gathering")
+            print(f"Token entropy/varentropy map: {self.hallu_handler.token_probs_map}")
             prefill_response = self._engage_parameter_gathering(messages)
             model_response = prefill_response.choices[0].message.content
 

model_server/src/core/hallucination.py

@@ -33,7 +33,7 @@ HALLUCINATION_THRESHOLD_DICT = {
     },
     MaskToken.PARAMETER_VALUE.value: {
         "entropy": 0.5437773168087006,
-        "varentropy": 2.413176417350769,
+        "varentropy": 0.49826696813106713,
     },
 }