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

Feature #2380

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

Fires on
0.45% of tokens
Mostly in
agent
Layer
after block 4

Where it fires

  • agent 71%
  • worksheet 18%
  • chat 5%
  • story 4%
  • python 2%

Makes the model say

else understand together printsreetultiplHey alone

Makes it not say

Oncei flny going lot there great

Strongest examples

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

  1. chat 2.84
    <|user_end|><|assistant_start|><|think_start|>7+1: 7+1=8, 8
  2. chat 2.73
    =7, 7⏎1+7: 1+7=8, 8
  3. worksheet 2.73
    =8, 8⏎1+7: 1+7=8, 8
  4. chat 2.65
    =7, 7⏎1+7: 1+7=8, 8
  5. worksheet 2.64
    +647: 1+7=8, 8
  6. worksheet 2.6
    97?⏎01+97: 1+7=8, 8
  7. worksheet 2.59
    1 plus 7?⏎1+7: 1+7=8, 8
  8. worksheet 2.57
    3417+0001: 7+1=8, 8
  9. worksheet 2.56
    914⏎47+01: 7+1=8, 8
  10. worksheet 2.55
    95⏎01+67: 1+7=8, 8
  11. worksheet 2.44
    , 676⏎7+1: 7+1=8, 8
  12. worksheet 2.4
    0087+8241: 7+1=8, 8