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

Feature #1477

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

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

Where it fires

  • worksheet 42%
  • agent 24%
  • chat 18%
  • python 12%
  • story 4%

Makes the model say

utilsoo noises shapescharsriesnums grass

Makes it not say

your= put not! foundP sad

Strongest examples

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

  1. chat 6.94
    <|user_end|><|assistant_start|><|think_start|>06+38: 6+8=14, 4
  2. chat 6.59
    <|think_start|>006+748: 6+8=14, 4
  3. worksheet 6.57
    +81658: 6+8=14, 4
  4. chat 6.56
    <|think_start|>106+078: 6+8=14, 4
  5. chat 6.55
    <|user_end|><|assistant_start|><|think_start|>06+38: 6+8=14, 4
  6. chat 6.3
    <|user_end|><|assistant_start|><|think_start|>78+56: 8+6=14, 4
  7. worksheet 6.29
    76+41148: 6+8=14, 4
  8. worksheet 6.22
    1697⏎6+8: 6+8=14, 1
  9. worksheet 6.22
    048+8576: 8+6=14, 4
  10. worksheet 6.17
    70⏎56+88: 6+8=14, 4
  11. worksheet 6.15
    36+90348: 6+8=14, 4
  12. chat 6.11
    <|user_end|><|assistant_start|><|think_start|>38+76: 8+6=14, 4