Skip to content
mini-lab
Log in Sign up
Features of prelude-2

Feature #3266

Fires on “ with” (not labeled yet: read the examples)

Fires on
0.50% of tokens
Mostly in
agent
Layer
after block 4

Where it fires

  • agent 94%
  • chat 4%
  • story 1%
  • python 1%
  • worksheet 0%

Makes the model say

def piece her it his")double the

Makes it not say

3 asksnediescase * +ful

Strongest examples

On text the dictionary never trained on. The underlined token is where it fires most.

  1. chat 2.69
    Juliet?<|user_end|><|assistant_start|>I'm sorry, I can't help with that. I only
  2. chat 2.63
    tonight?<|user_end|><|assistant_start|>I'm sorry, I can't help with that. I only
  3. chat 2.61
    tonight?<|user_end|><|assistant_start|>I'm sorry, I can't help with that. I only
  4. chat 2.6
    invest in?<|user_end|><|assistant_start|>I'm sorry, I can't help with that. I only
  5. chat 2.56
    NA stand for?<|user_end|><|assistant_start|>I'm sorry, I can't help with that. I only
  6. chat 2.53
    enya?<|user_end|><|assistant_start|>I'm sorry, I can't help with that. I only
  7. chat 2.52
    in the world?<|user_end|><|assistant_start|>I'm sorry, I can't help with that. I only
  8. chat 2.52
    <|user_end|><|assistant_start|>I'm sorry, I can't help with that. I only
  9. chat 2.52
    kyo?<|user_end|><|assistant_start|>I'm sorry, I can't help with that. I only
  10. chat 2.52
    Canada?<|user_end|><|assistant_start|>I'm sorry, I can't help with that. I only
  11. chat 2.48
    tonight?<|user_end|><|assistant_start|>I'm sorry, I can't help with that. I only
  12. chat 2.47
    NA stand for?<|user_end|><|assistant_start|>I'm sorry, I can't help with that. I only