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

Feature #30

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

Fires on
0.92% of tokens
Mostly in
worksheet
Layer
after block 4

Where it fires

  • worksheet 66%
  • chat 27%
  • agent 5%
  • story 1%
  • python 1%

Makes the model say

=")$} sum see<|assistant_end|>So

Makes it not say

( tried tookat much cubecube three

Strongest examples

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

  1. worksheet 4.83
    1⏎00+97: 0+7+1=8,
  2. worksheet 4.56
    , 0⏎0+3: 0+3+1=4,
  3. worksheet 4.44
    3: 7+3+1=11,
  4. worksheet 4.44
    3⏎01+63: 1+3+1=5,
  5. worksheet 4.39
    ⏎897+083: 7+3+1=11,
  6. chat 4.39
    7, 7⏎0+3: 0+3+1=4,
  7. worksheet 4.38
    ⏎23+08: 3+8+1=12,
  8. worksheet 4.35
    39⏎47+72: 7+2+1=10,
  9. worksheet 4.33
    ⏎481+927: 1+7+1=9,
  10. chat 4.31
    0⏎00+97: 0+7+1=8,
  11. chat 4.3
    3⏎07+50: 7+0+1=8,
  12. chat 4.26
    4⏎00+88: 0+8+1=9,