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

Feature #28

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

Fires on
0.36% of tokens
Mostly in
chat
Layer
after block 4

Where it fires

  • chat 41%
  • story 33%
  • agent 20%
  • python 6%
  • worksheet 1%

Makes the model say

,." ideaid can.min!"

Makes it not say

unexpected Fluffy# itsine$ plus not

Strongest examples

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

  1. chat 6.08
    yo?<|user_end|><|assistant_start|>I'm sorry, I can't help with that. I only know
  2. chat 5.98
    est in?<|user_end|><|assistant_start|>I'm sorry, I can't help with that. I only know
  3. chat 5.98
    ada?<|user_end|><|assistant_start|>I'm sorry, I can't help with that. I only know
  4. chat 5.95
    uliet?<|user_end|><|assistant_start|>I'm sorry, I can't help with that. I only know
  5. chat 5.94
    ch fast?<|user_end|><|assistant_start|>I'm sorry, I can't help with that. I only know
  6. chat 5.83
    I end?<|user_end|><|assistant_start|>I'm sorry, I can't help with that. I only know
  7. chat 5.83
    stand for?<|user_end|><|assistant_start|>I'm sorry, I can't help with that. I only know
  8. chat 5.83
    Cup?<|user_end|><|assistant_start|>I'm sorry, I can't help with that. I only know
  9. chat 5.8
    night?<|user_end|><|assistant_start|>I'm sorry, I can't help with that. I only know
  10. chat 5.75
    ork today?<|user_end|><|assistant_start|>I'm sorry, I can't help with that. I only know
  11. chat 5.72
    a cat?<|user_end|><|assistant_start|>I'm sorry, I can't help with that. I only know
  12. chat 5.72
    <|user_end|><|assistant_start|>I'm sorry, I can't help with that. I only know