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

Feature #2478

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

Fires on
0.34% of tokens
Mostly in
chat
Layer
after block 4

Where it fires

  • chat 66%
  • story 15%
  • agent 10%
  • worksheet 9%
  • python 1%

Makes the model say

<|arg|> knows wanted do()`.Olength't

Makes it not say

=3 number9 eyes# plus4

Strongest examples

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

  1. chat 4.87
    <|user_end|><|assistant_start|>I'm sorry, I can't help with that. I only know how to tell
  2. chat 4.81
    <|user_end|><|assistant_start|>I'm sorry, I can't help with that. I only know how to tell
  3. chat 4.8
    <|user_end|><|assistant_start|>I'm sorry, I can't help with that. I only know how to tell
  4. chat 4.79
    <|user_end|><|assistant_start|>I'm sorry, I can't help with that. I only know how to tell
  5. chat 4.77
    <|user_end|><|assistant_start|>I'm sorry, I can't help with that. I only know how to tell
  6. chat 4.76
    <|user_end|><|assistant_start|>I'm sorry, I can't help with that. I only know how to tell
  7. chat 4.75
    <|user_end|><|assistant_start|>I'm sorry, I can't help with that. I only know how to tell
  8. chat 4.75
    <|user_end|><|assistant_start|>I'm sorry, I can't help with that. I only know how to tell
  9. chat 4.74
    <|user_end|><|assistant_start|>I'm sorry, I can't help with that. I only know how to tell
  10. chat 4.74
    <|user_end|><|assistant_start|>I'm sorry, I can't help with that. I only know how to tell
  11. chat 4.74
    <|user_end|><|assistant_start|>I'm sorry, I can't help with that. I only know how to tell
  12. chat 4.7
    <|user_end|><|assistant_start|>I'm sorry, I can't help with that. I only know how to tell