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

Feature #1965

Fires on “3” (not labeled yet: read the examples)

Fires on
1.04% of tokens
Mostly in
agent
Layer
after block 4

Where it fires

  • agent 36%
  • chat 27%
  • worksheet 27%
  • python 6%
  • story 3%

Makes the model say

],! + for all p new short

Makes it not say

Theysum TheyloldString hitselvesva

Strongest examples

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

  1. python 3.23
    3⏎print("all
  2. worksheet 3.23
    3070 =
  3. chat 3.23
    36 + 2
  4. worksheet 3.23
    3⏎210
  5. agent 3.23
    3, in <module
  6. python 3.23
    3) == 6
  7. worksheet 3.23
    3586.
  8. chat 3.23
    35.<|assistant_end|><|user_start|>
  9. worksheet 3.23
    3 = 56
  10. worksheet 3.12
    0+2=2
  11. chat 3.12
    0 and 0.
  12. agent 3.12
    0]⏎```<|assistant_end|>