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

Feature #261

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

Fires on
0.81% of tokens
Mostly in
agent
Layer
after block 4

Where it fires

  • agent 58%
  • chat 22%
  • story 17%
  • python 2%
  • worksheet 1%

Makes the model say

way'tit6httpsute(" shouted

Makes it not say

⏎ then<|bos|>⏎⏎: togetherSoret

Strongest examples

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

  1. chat 2.26
    tiny model: I can only tell short stories
  2. chat 2.16
    tonight<|user_end|><|assistant_start|>Sorry, I'm a tiny model: I can only tell short stories
  3. chat 2.15
    of a cat<|user_end|><|assistant_start|>Sorry, I'm a tiny model: I can only tell short stories
  4. chat 2.15
    the moon?<|user_end|><|assistant_start|>Sorry, I'm a tiny model: I can only tell short stories
  5. chat 2.14
    dney?<|user_end|><|assistant_start|>Sorry, I'm a tiny model: I can only tell short stories
  6. chat 2.14
    good morning.<|user_end|><|assistant_start|>Sorry, I'm a tiny model: I can only tell short stories
  7. chat 2.11
    axes?<|user_end|><|assistant_start|>Sorry, I'm a tiny model: I can only tell short stories
  8. chat 2.1
    night?<|user_end|><|assistant_start|>Sorry, I'm a tiny model: I can only tell short stories
  9. agent 2.09
    test_lib.py`.<|assistant_end|><|user_start|>"help"<|user_end|><|assistant_start|>I can list, read,
  10. agent 2.09
    agent harness.<|system_end|><|user_start|>How can you help me?<|user_end|><|assistant_start|>I can list, read,
  11. chat 2.09
    for me.<|user_end|><|assistant_start|>Sorry, I'm a tiny model: I can only tell short stories
  12. chat 2.08
    stand for?<|user_end|><|assistant_start|>Sorry, I'm a tiny model: I can only tell short stories