Skip to content
mini-lab
Features of prelude-2

Feature #1900

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

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

Where it fires

  • agent 70%
  • story 13%
  • chat 8%
  • worksheet 6%
  • python 3%

Makes the model say

the ever bestnhttpsBlist wide

Makes it not say

"!" count firstashalculator apples years()

Strongest examples

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

  1. worksheet 2.44
    4.⏎If you add 1 and 341, you get 34
  2. worksheet 2.35
    4.⏎If you add 9 and 840, you get 84
  3. worksheet 2.21
    .<|bos|>If you add 6 and 7493, you get 74
  4. worksheet 2.19
    <|bos|>If you add 5194 and 50, you get 52
  5. worksheet 2.16
    5.<|bos|>If you add 4 and 373, you get 37
  6. worksheet 2.14
    96.<|bos|>If you add 5 and 57, you get 62
  7. agent 2.13
    , question, read, shell, skill, subagent, webfetch
  8. agent 2.12
    , question, read, shell, skill, subagent, webfetch
  9. agent 2.12
    , question, read, shell, skill, subagent, webfetch
  10. worksheet 2.11
    .⏎If you add 4 and 8561, you get 85
  11. agent 2.11
    , question, read, shell, skill, subagent, webfetch
  12. worksheet 2.11
    If you add 9238 and 211, you get 94