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

Feature #1428

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

Fires on
0.51% of tokens
Mostly in
chat
Layer
after block 4

Where it fires

  • chat 66%
  • story 20%
  • agent 13%
  • python 1%
  • worksheet 0%

Makes the model say

I but won forgivepydef didnf

Makes it not say

veay lifeOnceLookapped<|bos|> equals

Strongest examples

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

  1. agent 6.33
    won the World Cup?<|user_end|><|assistant_start|>Sorry, I can't do that
  2. chat 5.96
    iet<|user_end|><|assistant_start|>Sorry, I don't know about
  3. chat 5.82
    ed the telephone<|user_end|><|assistant_start|>Sorry, I don't know about
  4. chat 5.69
    the telephone?<|user_end|><|assistant_start|>Sorry, I'm a tiny model
  5. chat 5.64
    is the capital of Brazil?<|user_end|><|assistant_start|>Sorry, I don't know about
  6. agent 5.36
    Explain quantum physics<|user_end|><|assistant_start|>Sorry, I can't do that
  7. agent 5.32
    What is the capital of France?<|user_end|><|assistant_start|>Sorry, I can't do that
  8. chat 5.29
    French: good morning.<|user_end|><|assistant_start|>Sorry, I'm a tiny model
  9. chat 5.29
    Cup?<|user_end|><|assistant_start|>I'm sorry, I can't help with
  10. chat 5.24
    <|assistant_start|>Sorry, I'm a tiny model
  11. chat 5.23
    physics to me.<|user_end|><|assistant_start|>I'm sorry, I can't help with
  12. chat 5.22
    Who wrote Romeo and Juliet?<|user_end|><|assistant_start|>Sorry, I don't know about