Skip to content
mini-lab
Features of prelude-2

Feature #3232

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

Fires on
0.25% of tokens
Mostly in
agent
Layer
after block 4

Where it fires

  • agent 40%
  • python 28%
  • story 25%
  • chat 5%
  • worksheet 2%

Makes the model say

happened⏎ years([ eyesassertpython =

Makes it not say

)) familyIt returnpyve add all

Strongest examples

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

  1. story 2.31
    She was scared. "
  2. story 2.31
    She mixed everything together and
  3. story 2.31
    She had a favorite cup
  4. story 2.31
    She had a pet rabbit
  5. story 2.31
    She would make all kind
  6. story 2.28
    that had disappeared a few days ago. Tim was very persistent and would not
  7. agent 2.27
    <|user_start|>The tests are failing, fix the bug<|user_end|><|assistant_start|><|tool_call_start|>glob<|arg|>pattern=**/test
  8. agent 2.19
    tests are failing, fix the bug please"<|user_end|><|assistant_start|><|tool_call_start|>glob<|arg|>pattern=**/test
  9. agent 2.19
    The tests fail. Can you fix them?<|user_end|><|assistant_start|><|tool_call_start|>glob<|arg|>pattern=**/test
  10. story 2.18
    the sky. He loved to see the world from up above.⏎One
  11. agent 2.18
    The tests fail. Can you fix them?<|user_end|><|assistant_start|><|tool_call_start|>glob<|arg|>pattern=**/test
  12. agent 2.17
    he tests fail. Can you fix them?<|user_end|><|assistant_start|><|tool_call_start|>glob<|arg|>pattern=**/test