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

Feature #3193

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

Fires on
0.77% of tokens
Mostly in
agent
Layer
after block 4

Where it fires

  • agent 30%
  • python 30%
  • story 24%
  • chat 13%
  • worksheet 2%

Makes the model say

and with]) madelastet one can

Makes it not say

)`<|tool_call_end|>=**/*.Fix knowBugohopencode

Strongest examples

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

  1. agent 2.94
    NUMBERS = [6, 0, 0, 8,
  2. agent 2.89
    double_all([5, 1, 6, 0, 1]) ==
  3. chat 2.84
    <|user_start|>I want a story!<|user_end|><|assistant_start|>Max was a nosy three-year-old.
  4. story 2.82
    with her doll. She worked very hard and wanted to be a grown-up soon.
  5. agent 2.81
    VALUES = [9, 8, 7, 2, 2]⏎
  6. chat 2.72
    <|system_end|><|user_start|>Tell me a story.<|user_end|><|assistant_start|>One day, a kangaroo named K
  7. story 2.68
    ured over to take a closer look.⏎The monkey was mesmerized by the
  8. story 2.6
    started to talk! The tree said, "I am the memory in the box
  9. python 2.6
    ([5, 2, 5, 8, 7],
  10. agent 2.58
    [3, 8, 6, 1,
  11. python 2.57
    ⏎>>> evens([3, 7, 3, 0, 8])⏎
  12. python 2.57
    , `tail([9, 4, 0, 6, 0])` returns