revert to less terrible version

app.py

@@ -67,15 +67,14 @@ def build_messages(
 ) -> List[Dict[str, str]]:
     """Transform Gradio history into chat template messages.
 
-    History is stored as (human_assistant, model_user) for display purposes.
-    We extract them in the right order for the model.
+    History is stored as (model_user, human_assistant) tuples.
     """
     messages: List[Dict[str, str]] = []
     if system_prompt.strip():
         messages.append({"role": "system", "content": system_prompt.strip()})
 
-    # Each tuple is (human_assistant, model_user) for display
-    for human_assistant, model_user in history:
+    # Each tuple is (model_user, human_assistant)
+    for model_user, human_assistant in history:
         if model_user:
             messages.append({"role": "user", "content": model_user})
         if human_assistant:
@@ -115,8 +114,8 @@ def is_verbatim_repetition(
     if new_text_normalized == system_prompt.strip().lower():
         return True
 
-    # Check against previous model user messages (second element in tuple now)
-    for _, model_user in history:
+    # Check against previous model user messages (first element in tuple)
+    for model_user, _ in history:
         if model_user and new_text_normalized == model_user.strip().lower():
             return True
 
@@ -201,7 +200,7 @@ def respond(
     - If history empty: Generate first user message (ignores assistant_message input)
     - If history exists: Add assistant response and generate next user turn
 
-    History format: (human_assistant, model_user) for correct display sides
+    History format: (model_user, human_assistant)
     """
 
     # First message generation - ignore any text in the assistant box
@@ -218,8 +217,8 @@ def respond(
             top_p=top_p,
         )
 
-        # Start conversation with first user message
-        chat_history = [(None, user_reply)]
+        # Start conversation with first user message (empty assistant slot)
+        chat_history = [(user_reply, None)]
         return chat_history, chat_history
 
     # Subsequent messages - require assistant response
@@ -231,8 +230,8 @@ def respond(
         return chat_history, chat_history
 
     # Update the last tuple with the assistant response
-    _, last_model_user = chat_history[-1]
-    chat_history[-1] = (assistant_message.strip(), last_model_user)
+    last_model_user, _ = chat_history[-1]
+    chat_history[-1] = (last_model_user, assistant_message.strip())
 
     # Build messages for next user turn generation
     messages = build_messages(system_prompt, chat_history)
@@ -246,8 +245,8 @@ def respond(
         top_p=top_p,
     )
 
-    # Add new model user message
-    chat_history.append((None, user_reply))
+    # Add new model user message (with empty assistant slot)
+    chat_history.append((user_reply, None))
 
     return chat_history, chat_history
 
@@ -299,7 +298,7 @@ with gr.Blocks(theme=gr.themes.Soft()) as demo:
         submit_btn = gr.Button("Generate", variant="primary")
         clear_btn = gr.Button("Clear")
 
-    state = gr.State([])  # chat history: List[Tuple[human_assistant, model_user]]
+    state = gr.State([])  # chat history: List[Tuple[model_user, human_assistant]]
 
     with gr.Accordion("Implementation Details", open=False):
         gr.Markdown(
@@ -341,4 +340,4 @@ with gr.Blocks(theme=gr.themes.Soft()) as demo:
     clear_btn.click(_clear, outputs=[state, system_box, chatbot, msg])
 
 if __name__ == "__main__":
-    demo.queue().launch()
+    demo.queue().launch()