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

Feature #2852

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

Fires on
0.61% of tokens
Mostly in
story
Layer
after block 4

Where it fires

  • story 46%
  • agent 32%
  • chat 12%
  • python 9%
  • worksheet 1%

Makes the model say

but. In n time But fletter

Makes it not say

moral " unexpected<|tool_call_start|>:ied⏎ SCORES

Strongest examples

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

  1. agent 2.48
    trained from scratch on a laptop. I can list, read, run, create, edit and
  2. agent 2.43
    <|assistant_end|><|user_start|>"help"<|user_end|><|assistant_start|>I can list, read, run, create, edit and
  3. agent 2.43
    Hey there"<|user_end|><|assistant_start|>Hey! I can list, read, run, create, edit and
  4. agent 2.42
    only knows small Python projects. I can list, read, run, create, edit and
  5. python 2.39
    popular.<|assistant_end|><|bos|># notes.txt⏎Remember to run the tests.⏎⏎
  6. agent 2.37
    can you help me?<|user_end|><|assistant_start|>I can list, read, run, create, edit and
  7. agent 2.34
    trained from scratch on a laptop. I can list, read, run, create, edit and
  8. agent 2.33
    trained from scratch on a laptop. I can list, read, run, create, edit and
  9. agent 2.33
    s. I can list, read, run, create, edit and
  10. story 2.33
    many things, like big
  11. chat 2.33
    many flowers does Tim have
  12. agent 2.32
    <|user_start|>yo<|user_end|><|assistant_start|>Hey! I can list, read, run, create, edit and