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

Feature #3197

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

Fires on
1.16% of tokens
Mostly in
chat
Layer
after block 4

Where it fires

  • chat 28%
  • python 25%
  • agent 23%
  • story 21%
  • worksheet 2%

Makes the model say

back! another =='s return shade friends

Makes it not say

but go>>>pe squared somethingBut1

Strongest examples

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

  1. agent 7.01
    Hello!"<|user_end|><|assistant_start|>Hi
  2. agent 7.01
    Hello, Eva!"
  3. chat 5.51
    Hi!<|user_end|><|assistant_start|>Hi
  4. chat 4.54
    ?<|assistant_end|><|user_start|>Hello!<|user_end|><|assistant_start|>Hello!
  5. python 4.42
    "⏎assert greet("Noa") == "Hello, Noa!"⏎print
  6. story 3.98
    , "Hello! Why are you ring
  7. agent 3.93
    6: assert greet("Mia") == "Hello, Mia!"⏎⏎
  8. python 3.91
    ⏎from helpers import welcome⏎⏎assert welcome("Tom") == "Hello, Tom!"⏎
  9. chat 3.89
    only tell short stories and add numbers.<|assistant_end|><|user_start|>Hi<|user_end|><|assistant_start|>Hello! How can I help
  10. agent 3.82
    return "Hello, " + "
  11. python 3.76
    ython") == "python"⏎assert welcome("Sam") == "Hello, Sam!"⏎
  12. agent 3.76
    <|user_end|><|assistant_start|>Hello! How can I help