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

Feature #2926

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

Fires on
0.30% of tokens
Mostly in
agent
Layer
after block 4

Where it fires

  • agent 74%
  • python 20%
  • story 3%
  • chat 2%
  • worksheet 1%

Makes the model say

couldn along allndAt walked editquiet

Makes it not say

.2."th91se4

Strongest examples

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

  1. agent 3.97
    only knows small Python projects. I can list, read, run, create,
  2. agent 3.94
    language model trained from scratch on a laptop. I can list, read, run, create,
  3. agent 3.76
    language model trained from scratch on a laptop. I can list, read, run, create,
  4. agent 3.73
    s. I can list, read, run, create,
  5. agent 3.72
    language model trained from scratch on a laptop. I can list, read, run, create,
  6. agent 3.68
    language model trained from scratch on a laptop. I can list, read, run, create,
  7. agent 3.63
    . I can list, read, run, create,
  8. agent 3.63
    . I can list, read, run, create,
  9. agent 3.59
    language model trained from scratch on a laptop. I can list, read, run, create,
  10. agent 3.59
    <|user_start|>How can you help me?<|user_end|><|assistant_start|>I can list, read, run, create,
  11. agent 3.57
    can list, read, run, create,
  12. agent 3.54
    language model trained from scratch on a laptop. I can list, read, run, create,