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

Feature #3218

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

Fires on
0.17% of tokens
Mostly in
worksheet
Layer
after block 4

Where it fires

  • worksheet 31%
  • story 29%
  • chat 21%
  • agent 16%
  • python 4%

Makes the model say

food st swim(cubed sher always

Makes it not say

nd behind1<|assistant_end|> Spot cloudsTim Jack

Strongest examples

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

  1. agent 2.36
    1: def power(
  2. worksheet 2.36
    17 + 2
  3. chat 2.36
    1, 11
  4. agent 2.36
    1-6⏎1
  5. worksheet 2.36
    1, 1⏎
  6. python 2.36
    1]⏎⏎⏎def
  7. worksheet 2.36
    1=2,
  8. agent 2.36
    1: def includes(
  9. agent 2.36
    1⏎[6,
  10. chat 2.36
    16.<|assistant_end|><|user_start|>
  11. python 2.36
    14, 2
  12. worksheet 2.36
    10, 0