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

Feature #1752

Fires on “8” (not labeled yet: read the examples)

Fires on
0.85% of tokens
Mostly in
chat
Layer
after block 4

Where it fires

  • chat 46%
  • worksheet 26%
  • agent 17%
  • story 10%
  • python 0%

Makes the model say

?" did you round does!"a red

Makes it not say

equalsOnce$"`.Thank InsteadlloMIT

Strongest examples

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

  1. chat 3.48
    <|system_end|><|user_start|>Hey, what's 8 plus 83
  2. chat 3.22
    <|system_end|><|user_start|>What's 204 +
  3. chat 3.13
    <|user_end|><|assistant_start|>Happy to help!<|assistant_end|><|user_start|>Hey, what's 421 plus
  4. chat 3.12
    <|system_end|><|user_start|>What's 78 + 3
  5. story 3.1
    saw the spell book and said, "Tom, what is this book?" Tom answered,
  6. chat 3.03
    "Sure!".<|system_end|><|user_start|>What is 277 +
  7. chat 3.01
    <|system_start|>Answer with the number only.<|system_end|><|user_start|>Hey, what's 90 plus 1
  8. chat 2.99
    ```<|assistant_end|><|bos|><|system_start|>tools: calculator<|system_end|><|user_start|>Hey, what's 71 plus 9
  9. story 2.97
    a flag and some people on it.⏎"Tom, what is that?" Lily asked.
  10. chat 2.97
    calculator<|system_end|><|system_start|>Do not use the calculator.<|system_end|><|user_start|>What is 87 plus 6
  11. chat 2.97
    ! Have a nice day.<|assistant_end|><|user_start|>Hey, what's 9 plus 4?
  12. chat 2.96
    <|assistant_end|><|bos|><|system_start|>Answer with the number only.<|system_end|><|user_start|>What is 832 plus