change handler

handler.py

@@ -36,9 +36,9 @@ class EndpointHandler:
         inputs = data.get("inputs", "")
         parameters = data.get("parameters", {})
         
-        # Prepare the conversation
+        # Prepare the conversation for text splitting task
         messages = [
-            {"role": "system", "content": "Eres un asistente conversacional amigable especializado en conversaciones tipo WhatsApp en español."},
+            {"role": "system", "content": "Split messages at natural breaks into JSON array. Common patterns: greeting+question, statement+question, topic+followup. Keep original words, only add logical splits."},
             {"role": "user", "content": inputs}
         ]
         
@@ -50,7 +50,7 @@ class EndpointHandler:
                 add_generation_prompt=True
             )
         else:
-            text = f"Usuario: {inputs}\nAsistente:"
+            text = f"Split messages at natural breaks into JSON array. Common patterns: greeting+question, statement+question, topic+followup. Keep original words, only add logical splits.\nUser: {inputs}\nAssistant:"
         
         # Tokenize
         model_inputs = self.tokenizer(text, return_tensors="pt").to(self.model.device)
@@ -59,11 +59,12 @@ class EndpointHandler:
         with torch.no_grad():
             outputs = self.model.generate(
                 **model_inputs,
-                max_new_tokens=parameters.get("max_new_tokens", 200),
-                temperature=parameters.get("temperature", 0.7),
+                max_new_tokens=parameters.get("max_new_tokens", 100),
+                temperature=parameters.get("temperature", 0.3),  # Lower for consistent JSON format
                 top_p=parameters.get("top_p", 0.9),
                 do_sample=True,
-                pad_token_id=self.tokenizer.eos_token_id
+                pad_token_id=self.tokenizer.eos_token_id,
+                repetition_penalty=1.1
             )
         
         # Decode response