Skip to content
mini-lab
Features of prelude-2

Feature #2452

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

Fires on
1.30% of tokens
Mostly in
story
Layer
after block 4

Where it fires

  • story 74%
  • chat 15%
  • agent 5%
  • python 4%
  • worksheet 2%

Makes the model say

her hisCan watch( him a([

Makes it not say

<|assistant_end|>$ FalseAnd and!. both

Strongest examples

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

  1. story 3.2
    the tree.<|bos|>One day, a little boy named Tim was playing in his room. He
  2. story 3.1
    2`.⏎<|bos|>One day, a little girl named Lily was playing in her room. She
  3. story 3.08
    now.<|assistant_end|><|bos|>One day, a little girl named Lucy was playing in her room. She
  4. story 2.99
    236.<|bos|>One day, a girl named Lily was playing in her room. She
  5. story 2.97
    5.<|assistant_end|><|bos|>One day, a little girl named Amy was playing in her room. She
  6. story 2.93
    today?<|assistant_end|><|bos|>One day, a little girl named Mia was playing in her room. She
  7. story 2.92
    ⏎```<|assistant_end|><|bos|>One day, a little dog named Spot was playing in the park. He
  8. story 2.86
    2.<|assistant_end|><|bos|>One day, a little girl named Mia was playing in her room. She
  9. story 2.84
    a time, there was a little boy named Tim. Tim was playing in his room when he
  10. chat 2.84
    <|user_end|><|assistant_start|>One day, a little dog named Pup was playing in the park.<|assistant_end|>
  11. story 2.83
    apon.<|bos|>One day, a cat named Bow was walking in the yard. B
  12. story 2.79
    very happy.<|assistant_end|><|bos|>One day, a boy named Tim was playing in his yard. He