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

Feature #2645

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

Fires on
0.62% of tokens
Mostly in
agent
Layer
after block 4

Where it fires

  • agent 58%
  • chat 35%
  • story 5%
  • python 1%
  • worksheet 0%

Makes the model say

is until was to Then will model listened

Makes it not say

branch Lilyase()`], rever Tom Sue

Strongest examples

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

  1. chat 2.66
    87.<|user_end|><|assistant_start|>The answer is 91
  2. chat 2.65
    <|assistant_start|>The answer is 6.
  3. chat 2.65
    <|assistant_start|>The answer is 41
  4. chat 2.65
    <|assistant_start|>The answer is 10
  5. chat 2.65
    <|assistant_start|>The answer is 76
  6. chat 2.65
    <|assistant_start|>The answer is 9.
  7. chat 2.65
    <|assistant_start|>The answer is 24
  8. chat 2.65
    <|assistant_start|>The answer is 86
  9. chat 2.62
    7.<|user_end|><|assistant_start|>The answer is 11
  10. chat 2.61
    35.<|user_end|><|assistant_start|>The answer is 92
  11. agent 2.6
    <|user_end|><|assistant_start|>Sorry, I can't do that: I'm a tiny model that only knows small
  12. chat 2.6
    3.<|user_end|><|assistant_start|>The answer is 10