rebuild

app.py

@@ -14,7 +14,7 @@ HF_TOKEN = os.environ.get("HF_TOKEN")
 login(token=HF_TOKEN)
 
 MAX_MAX_NEW_TOKENS = 2048
-DEFAULT_MAX_NEW_TOKENS = 256  # smaller default to save memory
+DEFAULT_MAX_NEW_TOKENS = 128  # smaller default to save memory
 MAX_INPUT_TOKEN_LENGTH = 4096
 
 css = """
@@ -178,7 +178,6 @@ def generate(
             }
         ] if subspaces_list else None,
         "streamer": streamer,
-        "repetition_penalty": 1.5,
         "do_sample": True
     }
 
@@ -222,8 +221,8 @@ def update_dropdown_choices(search_text):
     filtered = filter_concepts(search_text)
     if not filtered or len(filtered) == 0:
         return gr.update(choices=[f"[New] {search_text}"], value=f"[New] {search_text}", interactive=True), gr.Textbox(
-        label="No matching existing concepts were found!", 
-        value="Good news! Based on the concept you provided, we will automatically generate a steering vector. Try it out by starting a chat!",
+        label="No matching existing topics were found!", 
+        value="Good news! Based on the topic you provided, we will automatically generate a steering vector. Try it out by starting a chat!",
         lines=3,
         interactive=False,
         visible=True,
@@ -240,43 +239,42 @@ with gr.Blocks(css=css, fill_height=True) as demo:
     # Remove default subspaces
     selected_subspaces = gr.State([])
     
-    with gr.Row(min_height=700):
+    with gr.Row(min_height=300):
         # Left side: bigger chat area
         with gr.Column(scale=7):
             chat_interface = gr.ChatInterface(
                 fn=generate,
-                title="Chat with a Concept Steering Model",
-                description="""Steer responses by selecting concepts on the right →\n\nWe intervene on Gemma-2-2B-it by adding steering vectors to the residual stream at layer 20.""",
+                title="Chat with a Topic Steering Model",
+                description="""Choose a topic you want the model to discuss on the right →\n\nWe intervene on Gemma-2-2B-it by adding steering vectors to the residual stream at layer 20.""",
                 type="messages",
                 additional_inputs=[selected_subspaces],
-                fill_height=True
             )
         
         # Right side: concept management
         with gr.Column(scale=3):
-            gr.Markdown("## Steer Model Responses")
-            gr.Markdown("Search and then select a concept to steer. The closest match will be automatically selected. If there is no match, our steering vector generator (a finetuned Gemma-2-2B) auto-steers for you!")
+            gr.Markdown("# Steer model responses")
+            gr.Markdown("Search and then select a topic you want the model to discuss. The closest match will be automatically selected. If there is no match, a finetuned Gemma-2-2B model auto-steers for you!")
             # Concept Search and Selection
             with gr.Group():
                 search_box = gr.Textbox(
-                    label="Search Concepts",
-                    placeholder="Find concepts to steer the model (e.g. 'time travel')",
+                    label="Search topics to steer",
+                    placeholder="Try: 'time travel'",
                     lines=2,
                 )
                 msg = gr.TextArea(visible=False)
                 concept_dropdown = gr.Dropdown(
-                    label="Select a concept to steer the model (Click to see more!)",
+                    label="Select a topic to steer the model (Click to see more!)",
                     interactive=True,
                     allow_custom_value=False,
                 )
                 concept_magnitude = gr.Slider(
-                    label="Steering Intensity",
+                    label="Steering intensity",
                     minimum=-5,
                     maximum=5,
                     step=0.1,
                     value=3,
                 )
-
+            
     # Wire up events
     # When search box changes, update dropdown AND trigger concept selection
     search_box.input(