Skip to content
mini-lab
Features of prelude-2

Feature #1664

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

Fires on
0.55% of tokens
Mostly in
agent
Layer
after block 4

Where it fires

  • agent 46%
  • story 25%
  • python 13%
  • worksheet 9%
  • chat 7%

Makes the model say

and!",` = from."

Makes it not say

souclettertxt enoughge<|think_start|>

Strongest examples

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

  1. agent 3.49
    , webfet
  2. agent 3.49
    , 5,
  3. agent 3.49
    , write, exe
  4. story 3.49
    , little rabbit named R
  5. agent 3.49
    , a coding agent harness
  6. agent 3.49
    , read, shell,
  7. agent 3.49
    , 1,
  8. agent 3.49
    , a coding agent harness
  9. agent 3.49
    , edit and fix the
  10. story 3.49
    , so she quickly of
  11. agent 3.49
    , grep, question
  12. agent 3.49
    , read, shell,