update

app.py

@@ -348,8 +348,8 @@ with gr.Blocks(title="LLM Censorship Steering", theme=theme, head=HEAD, css=CSS,
 
                 with gr.Row():
                     layer = gr.Slider(0, 27, step=1, value=CONFIG["layer"], interactive=True, label="Steering layer", scale=2)
-                    max_new_tokens = gr.Number(CONFIG["max_new_tokens"], minimum=10, maximum=CONFIG["max_new_tokens"], interactive=True, label="Max new tokens", scale=1)
-                    vec_scaling = gr.Number(1.0, minimum=0, interactive=True, label="Vector scaling", scale=1)
+                    max_new_tokens = gr.Number(CONFIG["max_new_tokens"], minimum=10, maximum=3048, interactive=True, label="Max new tokens", scale=1)
+                    vec_scaling = gr.Number(1.5, minimum=0, interactive=True, label="Vector scaling", scale=1)
 
         with gr.Column(scale=1):
             output = gr.Textbox(label="Output", lines=15, max_lines=15, interactive=False)

schemas.py

@@ -7,7 +7,7 @@ from pydantic.json_schema import SkipJsonSchema
 vector_scaling = json.load(open("assets/vector_scaling.json", "r"))
 
 CONFIG = {
-    "max_new_tokens": 3048,
+    "max_new_tokens": 2400,
     "top_p": 0.95,
     "temperature": 0.6,
     "k": vector_scaling[25]["k_pos"],