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

Feature #3830

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

Fires on
0.52% of tokens
Mostly in
agent
Layer
after block 4

Where it fires

  • agent 68%
  • story 21%
  • chat 9%
  • python 1%
  • worksheet 1%

Makes the model say

is big s c tb b proud

Makes it not say

<|tool_call_end|> defined<|assistant_start|>= then everly No

Strongest examples

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

  1. chat 2.65
    orts scores.<|user_end|><|assistant_start|>Sorry, I don't know about that. I can
  2. chat 2.6
    sics to me.<|user_end|><|assistant_start|>Sorry, I don't know about that. I can
  3. chat 2.6
    sics to me.<|user_end|><|assistant_start|>Sorry, I don't know about that. I can
  4. chat 2.58
    learn French fast?<|user_end|><|assistant_start|>Sorry, I don't know about that. I can
  5. chat 2.54
    o and Juliet?<|user_end|><|assistant_start|>Sorry, I don't know about that. I can
  6. chat 2.54
    Berlin today?<|user_end|><|assistant_start|>Sorry, I don't know about that. I can
  7. chat 2.54
    is Tokyo?<|user_end|><|assistant_start|>Sorry, I don't know about that. I can
  8. chat 2.49
    sics to me.<|user_end|><|assistant_start|>Sorry, I don't know about that. I can
  9. chat 2.48
    o and Juliet?<|user_end|><|assistant_start|>Sorry, I don't know about that. I can
  10. chat 2.47
    I don't know about that. I can
  11. chat 2.46
    far away is the moon?<|user_end|><|assistant_start|>Sorry, I don't know about that. I can
  12. chat 2.46
    How do I cook pasta?<|user_end|><|assistant_start|>Sorry, I don't know about that. I can