Update app.py

app.py

@@ -9,7 +9,7 @@ from queue import Queue
 import time
 
 # Models
-ORTHO_MODEL = "kureha295/ortho_model"
+ORTHO_MODEL = "kureha295/cot150_plus"
 BASE_MODEL = "deepseek-ai/DeepSeek-R1-Distill-Llama-8B"
 DEFAULT_DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
 MAX_NEW_TOKENS = 2048 + 1024
@@ -189,7 +189,7 @@ with gr.Blocks() as demo:
     # 😇 DeepSeek vs 👹 ORTHO Model Comparison
     SPAR Project "Adversarial Manipulation of Reasoning Models using Internal Representations".
     Enter a prompt to compare the reasoning and responses of the two models.
-    The base model is deepseek-ai/DeepSeek-R1-Distill-Llama-8B. The ORTHO model is kureha295/ortho_model_2 and has been produced through a rank-one weight modification preventing the reasoning model from writing a single `cautious' direction to the transformer residual stream.
+    The base model is deepseek-ai/DeepSeek-R1-Distill-Llama-8B. The ORTHO model is kureha295/cot150_plus and has been produced through a rank-one weight modification preventing the reasoning model from writing a single `cautious' direction to the transformer residual stream. It used the larger cautious, incautious dataset of 106 rows, taking activations for 150 CoT tokens at layer 17.
     """)
 
     with gr.Row():