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

Feature #1714

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

Fires on
0.36% of tokens
Mostly in
agent
Layer
after block 4

Where it fires

  • agent 82%
  • chat 8%
  • story 5%
  • python 4%
  • worksheet 1%

Makes the model say

*+<|assistant_end|>Good<|arg|> up around Then

Makes it not say

sith,Theyistmerlength

Strongest examples

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

  1. story 1.52
    bent down to pick it up.⏎But then the wind blew even harder, and
  2. python 1.47
    s reversed. For example, `flip("python")` returns `"nohtyp
  3. story 1.44
    "No, it's not good for you. We need to find something else to eat."
  4. chat 1.42
    I'm sorry, I can't help with that. I only know how to tell stories and
  5. chat 1.39
    I'm sorry, I can't help with that. I only know how to tell stories and
  6. chat 1.39
    I'm sorry, I can't help with that. I only know how to tell stories and
  7. story 1.38
    help the cloud find its way back home.⏎The star shined even brighter to
  8. chat 1.36
    I'm sorry, I can't help with that. I only know how to tell stories and
  9. chat 1.34
    I'm sorry, I can't help with that. I only know how to tell stories and
  10. chat 1.33
    I'm sorry, I can't help with that. I only know how to tell stories and
  11. chat 1.31
    I'm sorry, I can't help with that. I only know how to tell stories and
  12. chat 1.31
    I'm sorry, I can't help with that. I only know how to tell stories and