Update app.py

app.py

@@ -16,22 +16,6 @@ if torch.backends.mps.is_available():
     model.to("mps")
 
 
-def process_history(history):
-    """Process chat history into the format expected by the model."""
-    processed_history = []
-    for user_msg, assistant_msg in history:
-        # Always add user message first, even if empty
-        processed_history.append({"role": "user", "content": [{"type": "text", "text": user_msg or ""}]})
-        # Always add assistant message, even if empty
-        processed_history.append({"role": "assistant", "content": [{"type": "text", "text": assistant_msg or ""}]})
-    return processed_history
-
-
-def process_new_user_message(message):
-    """Process a new user message into the format expected by the model."""
-    return [{"type": "text", "text": message}]
-
-
 def respond(
     user_message,
     history: list[tuple[str, str]],
@@ -45,9 +29,10 @@ def respond(
     if system_message:
         messages.append({"role": "system", "content": system_message})
     
-    # Process the conversation history
-    if history:
-        messages.extend(process_history(history))
+    # # Process the conversation history
+    # No need for history functionality for this reasoning model
+    # if history:
+    #     messages.extend(process_history(history))
     
     # Add the new user message
     messages.append({"role": "user", "content": user_message})
@@ -75,7 +60,6 @@ def respond(
         temperature=temperature,
         top_p=top_p,
         top_k=64,  # Recommended Gemma-3 setting
-        # do_sample=True,
     )
     
     thread = threading.Thread(target=model.generate, kwargs=generate_kwargs)
@@ -88,7 +72,6 @@ def respond(
         # Tags
         start_tag = "<start_working_out>"
         sol_start = "<SOLUTION>"
-        sol_end = "</SOLUTION>"
 
         thinking = ""
         final_answer = ""
@@ -106,12 +89,7 @@ def respond(
         # Extract "Final answer" section: everything after <SOLUTION>
         if sol_start in output:
             sol_start_idx = output.find(sol_start) + len(sol_start)
-            # If </SOLUTION> is present, stop at it
-            if sol_end in output:
-                sol_end_idx = output.find(sol_end)
-                final_answer = output[sol_start_idx:sol_end_idx].strip()
-            else:
-                final_answer = output[sol_start_idx:].strip()
+            final_answer = output[sol_start_idx:].strip()
 
         # Build formatted output
         formatted_output = ""
@@ -126,10 +104,6 @@ def respond(
 
         yield formatted_output
 
-        # If </SOLUTION> is found, end the response
-        if sol_end in output:
-            break
-
 
 """
 For information on how to customize the ChatInterface, peruse the gradio docs: https://www.gradio.app/docs/chatinterface
@@ -165,6 +139,13 @@ demo = gr.ChatInterface(
             {"left": "$", "right": "$", "display": False}
         ],
     ),
+    title="Demo: Finetuned Gemma 3 1B (Oysiyl/gemma-3-1B-GRPO) on GSM8K with GRPO",
+    description=(
+        "This is a demo of a finetuned Gemma 3 1B model ([Oysiyl/gemma-3-1B-GRPO](https://huggingface.co/Oysiyl/gemma-3-1B-GRPO)) "
+        "on the [openai/gsm8k](https://huggingface.co/datasets/openai/gsm8k) dataset using the GRPO technique. "
+        "Finetuning and reasoning approach inspired by [Unsloth's notebook](https://docs.unsloth.ai/basics/reasoning-grpo-and-rl/tutorial-train-your-own-reasoning-model-with-grpo). "
+        "This demo does not support conversation history, as the GSM8K dataset consists of single-turn questions."
+    ),
 )