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

Feature #1745

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

Fires on
0.42% of tokens
Mostly in
worksheet
Layer
after block 4

Where it fires

  • worksheet 59%
  • chat 32%
  • story 4%
  • agent 3%
  • python 2%

Makes the model say

,"s we off` onh

Makes it not say

atlioldStringctome<|user_end|>lelet

Strongest examples

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

  1. chat 7.09
    , 45⏎1+0: 1+0=1, 14
  2. chat 7.05
    , 98⏎0+1: 0+1=1, 19
  3. worksheet 6.77
    , 23⏎1+0: 1+0=1, 12
  4. worksheet 6.76
    , 49⏎1+0: 1+0=1, 14
  5. chat 6.76
    6, 6⏎1+0: 1+0=1, 16
  6. chat 6.75
    , 84⏎1+0: 1+0=1, 18
  7. worksheet 6.75
    340⏎1+0: 1+0=1, 13
  8. worksheet 6.68
    3809⏎0+1: 0+1=1, 13
  9. worksheet 6.67
    , 96⏎1+0: 1+0=1, 19
  10. chat 6.65
    , 29⏎1+0: 1+0=1, 12
  11. worksheet 6.65
    962⏎1+0: 1+0=1, 19
  12. chat 6.63
    , 39⏎1+0: 1+0=1, 13