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

Feature #3606

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

Fires on
0.40% of tokens
Mostly in
worksheet
Layer
after block 4

Where it fires

  • worksheet 52%
  • chat 26%
  • agent 11%
  • story 5%
  • python 5%

Makes the model say

1 divisibleion python by example()` that

Makes it not say

u equals⏎ makes isra were"

Strongest examples

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

  1. chat 5.33
    9, 9⏎8+3: 8+3=11, 1
  2. chat 5.01
    4, 4⏎3+8: 3+8=11, 1
  3. worksheet 4.98
    ⏎288+053: 8+3=11, 1
  4. chat 4.98
    3+38: 3+8=11, 1
  5. chat 4.93
    , 88⏎8+3: 8+3=11, 1
  6. worksheet 4.9
    48+85503: 8+3=11, 1
  7. worksheet 4.88
    08+00223: 8+3=11, 1
  8. chat 4.85
    8: 3+8=11, 1
  9. chat 4.83
    is?<|user_end|><|assistant_start|><|think_start|>8+3: 8+3=11, 1
  10. worksheet 4.8
    ⏎008+593: 8+3=11, 1
  11. worksheet 4.78
    13+98088: 3+8=11, 1
  12. worksheet 4.66
    ⏎433+078: 3+8=11, 1