Skip to content
mini-lab
Features of prelude-2

Feature #1918

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

Fires on
0.59% of tokens
Mostly in
worksheet
Layer
after block 4

Where it fires

  • worksheet 73%
  • chat 23%
  • agent 2%
  • python 1%
  • story 1%

Makes the model say

+$⏎6 < shells balls promise

Makes it not say

w rur when did do bear function

Strongest examples

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

  1. worksheet 4.6
    9⏎9701+0006: 1+6=7
  2. worksheet 4.51
    9⏎3301+1014: 1+4=5
  3. chat 4.5
    0?<|user_end|><|assistant_start|><|think_start|>001+990: 1+0=1
  4. worksheet 4.49
    + 4?⏎781+004: 1+4=5
  5. worksheet 4.48
    0⏎0501+9769: 1+9+1
  6. chat 4.45
    31<|user_end|><|assistant_start|><|think_start|>071+931: 1+1=2
  7. worksheet 4.39
    0003: 0+3=3
  8. worksheet 4.37
    9.⏎1021+4089: 1+9=1
  9. worksheet 4.35
    7.⏎0991+8807: 1+7=8
  10. chat 4.33
    that?<|user_end|><|assistant_start|><|think_start|>101+010: 1+0=1
  11. worksheet 4.29
    6?⏎9121+0006: 1+6=7
  12. chat 4.28
    8?<|user_end|><|assistant_start|><|think_start|>091+138: 1+8=9