Skip to content
mini-lab
Features of prelude-2

Feature #3631

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

Fires on
1.20% of tokens
Mostly in
worksheet
Layer
after block 4

Where it fires

  • worksheet 55%
  • chat 34%
  • python 9%
  • agent 1%
  • story 1%

Makes the model say

='s<|assistant_end|>/So tests)`. filled

Makes it not say

to0One helpigx everyone1

Strongest examples

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

  1. worksheet 3.46
    8+1=14, 470⏎0+1: 0+
  2. chat 3.46
    6+7=13, 38⏎4+1: 4+
  3. chat 3.43
    : 9+9=18, 8⏎5+1: 5+
  4. worksheet 3.41
    9+7=16, 683⏎0+1: 0+
  5. worksheet 3.4
    0+7=7, 735⏎0+1: 0+
  6. worksheet 3.4
    +1+1=2, 284⏎0+1: 0+
  7. chat 3.38
    3: 5+3=8, 8⏎6+5: 6+
  8. chat 3.38
    +7+1=17, 72⏎2+1: 2+
  9. worksheet 3.37
    1+0=1, 1854⏎5+0: 5+
  10. worksheet 3.35
    : 1+8=9, 97⏎0+7: 0+
  11. chat 3.34
    : 2+0=2, 29⏎1+0: 1+
  12. chat 3.34
    : 8+0=8, 84⏎5+0: 5+