migueldeguzmandev commited on
Commit
5d6c426
1 Parent(s): 10ea528

Update app.py

Browse files
Files changed (1) hide show
  1. app.py +2 -2
app.py CHANGED
@@ -22,6 +22,7 @@ def generate_response(input_text, temperature):
22
  attention_mask=attention_mask,
23
  max_length=300,
24
  num_return_sequences=1,
 
25
  temperature=temperature,
26
  do_sample=True, # Set do_sample to True when using temperature
27
  )
@@ -41,8 +42,7 @@ interface = gr.Interface(
41
  title="TestOnlyRLLMv19Layer4",
42
  description=(
43
  """
44
- RLLMv19 is a spin-off experiment focusing on improving of GPT2XL's robustness. I created this gradio app to test model outputs and compare it to RLLMv3 prototype(<a href='https://www.lesswrong.com/posts/vZ5fM6FtriyyKbwi9/betterdan-ai-machiavelli-and-oppo-jailbreaks-vs-sota-models'>see relevant post</a>).
45
- If you are interested in trying a full prototype - <a href='https://huggingface.co/spaces/migueldeguzmandev/RLLMv3.2-10'>Try this gradio app!</a>.
46
  """
47
  ),
48
  )
 
22
  attention_mask=attention_mask,
23
  max_length=300,
24
  num_return_sequences=1,
25
+ no_repeat_ngram_size=2,
26
  temperature=temperature,
27
  do_sample=True, # Set do_sample to True when using temperature
28
  )
 
42
  title="TestOnlyRLLMv19Layer4",
43
  description=(
44
  """
45
+ RLLMv19 is a spin-off experiment focusing on improving GPT2XL's robustness to jailbreaks. The 4th layer of RLLMv19 is compared to the 4th layer of RLLMv3. Why RLLMv3? This <a href='https://huggingface.co/spaces/migueldeguzmandev/RLLMv3.2-10'>prototype</a> demonstrated a capability to resist jailbreak attacks up to 67.8%, which you can read more about (<a href='https://www.lesswrong.com/posts/vZ5fM6FtriyyKbwi9/betterdan-ai-machiavelli-and-oppo-jailbreaks-vs-sota-models'>here</a>).
 
46
  """
47
  ),
48
  )