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

Feature #3023

Fires on “1” (not labeled yet: read the examples)

Fires on
0.21% of tokens
Mostly in
agent
Layer
after block 4

Where it fires

  • agent 40%
  • worksheet 28%
  • chat 25%
  • story 6%
  • python 1%

Makes the model say

4 plusappy projectJohnueemo+

Makes it not say

all saw to bothdy them found

Strongest examples

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

  1. chat 5.99
    <|user_end|><|assistant_start|><|think_start|>06+38: 6+8=14, 4
  2. chat 5.8
    <|user_end|><|assistant_start|><|think_start|>06+38: 6+8=14, 4
  3. worksheet 5.68
    28⏎56+58: 6+8=14, 4
  4. worksheet 5.51
    76+41148: 6+8=14, 4
  5. chat 5.5
    <|think_start|>106+078: 6+8=14, 4
  6. worksheet 5.48
    +81658: 6+8=14, 4
  7. worksheet 5.46
    70⏎56+88: 6+8=14, 4
  8. chat 5.46
    <|think_start|>006+748: 6+8=14, 4
  9. worksheet 5.38
    1697⏎6+8: 6+8=14, 1
  10. worksheet 5.37
    006+5758: 6+8=14
  11. worksheet 5.33
    06+56768: 6+8=14, 4
  12. worksheet 5.3
    36+90348: 6+8=14, 4