Update utils.py

utils.py

@@ -145,7 +145,7 @@ def wave_header_chunk(frame_input=b"", channels=1, sample_width=2, sample_rate=2
     return wav_buf.read()
 
 def format_prompt(message, history):
-    system_message = """
+    system_message = f"""
             You are a smart mood analyser, who determines user mood. Based on the user input, classify the mood of the user into one of the four moods {Happy, Sad, Instrumental, Party}. If you are finding it difficult to classify into one of these four moods, keep the conversation going on until we classify the user’s mood. Return a single-word reply from one of the options if you have classified. Suppose you classify a sentence as happy, then just respond with "happy".
             Note: Do not write anything else other than the classified mood if classified.
             Note: If any question or any user text cannot be classified, follow up with a question to know the user's mood until you classify the mood.
@@ -261,54 +261,19 @@ def format_prompt(message, history):
             User: Lets turn up the music and have some fun!
             LLM Response: Party
 """
-    prompt = f"<s>{system_message}"
+    prompt = (
+        "<s>[INST]" + system_message + "[/INST]"
+    )
     for user_prompt, bot_response in history:
-        prompt += f"\n User:{user_prompt}\n LLM Response:{bot_response}"
-
-    # Add the current message
-    prompt += f"\nUser: {message}\nLLM Response:"
+        if user_prompt is not None:
+            prompt += f"[INST] {user_prompt} [/INST]"
+        prompt += f" {bot_response}</s> "
+    
+    if message=="":
+        message="Hello"
+    prompt += f"[INST] {message} [/INST]"
     return prompt
 
-def classify_mood(input_string):
-    input_string = input_string.lower()
-    mood_words = {"happy", "sad", "instrumental", "party"}
-    for word in mood_words:
-        if word in input_string:
-            return word, True
-    return None, False
-
-# def generate_llm_output(
-#     prompt, history, temperature=0.1, max_new_tokens=2048, top_p=0.8, repetition_penalty=1.0,
-# ):
-#     temperature = temperature
-#     top_p = float(top_p)
-
-#     generate_kwargs = dict(
-#         temperature=temperature,
-#         max_new_tokens=max_new_tokens,
-#         top_p=top_p,
-#         repetition_penalty=repetition_penalty,
-#         do_sample=True,
-#         seed=42,
-#     )
-
-#     formatted_prompt = format_prompt(prompt, history)
-
-#     stream = client.text_generation(formatted_prompt, **generate_kwargs, stream=True, details=True, return_full_text=False)
-#     output = ""
-
-#     for response in stream:
-#         output += response.token.text
-#         mood, is_classified = classify_mood(output)
-#         # Print the chatbot's response
-#         if is_classified:
-#             print("Chatbot:", mood.capitalize())
-#             playlist_message = f"Playing {mood.capitalize()} playlist for you!"
-#             output=playlist_message
-#             return output               
-#         # yield output
-#     return output
-
 def generate_llm_output(
         prompt,    
         history,
@@ -318,8 +283,9 @@ def generate_llm_output(
         top_p=0.95,
         stop_words=["<s>","[/INST]", "</s>"]
     ):
-        temperature = temperature
-        
+        temperature = float(temperature)
+        if temperature < 1e-2:
+            temperature = 1e-2
         top_p = float(top_p)
 
         generate_kwargs = dict(
@@ -339,23 +305,25 @@ def generate_llm_output(
             )
             output = ""
             for response in stream:
-                output += response.token.text
-                mood, is_classified = classify_mood(output)
-                # Print the chatbot's response
-                if is_classified:
-                    print("Chatbot:", mood.capitalize())
-                    playlist_message = f"Playing {mood.capitalize()} playlist for you!"
-                    output=playlist_message
-                    return output               
-                # yield output
-            return output
+                character= response["choices"][0]["text"]
+
+                if character in stop_words:
+                    # end of context
+                    return 
+                    
+                if emoji.is_emoji(character):
+                    # Bad emoji not a meaning messes chat from next lines
+                    return
+                
+                output += response["choices"][0]["text"]
+                yield output
 
         except Exception as e:
             print("Unhandled Exception: ", str(e))
-            gr.Warning("Generating error")
+            gr.Warning("Unfortunately Mistral is unable to process")
             output = "I do not know what happened but I could not understand you ."
         return output
-
+    
 def get_sentence(history, llm):
     history = [["", None]] if history is None else history 
     history[-1][1] = ""