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

Feature #3909

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

Fires on
0.99% of tokens
Mostly in
agent
Layer
after block 4

Where it fires

  • agent 33%
  • worksheet 25%
  • python 21%
  • chat 19%
  • story 2%

Makes the model say

([ number`_only man angry turns

Makes it not say

<|assistant_end|>itedortHer6 going had5

Strongest examples

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

  1. worksheet 13.05
    is 330.<|assistant_end|><|bos|>Mia has 434 apples. Mia gets
  2. worksheet 12.74
    6 = 358.⏎Sue has 474 apples. Sue gets
  3. worksheet 12.74
    95100.⏎Mia has 823 apples. Mia gets
  4. worksheet 12.53
    154<|assistant_end|><|bos|>Tom has 24320 shells. Tom gets
  5. worksheet 12.49
    ! Have a nice day.<|assistant_end|><|bos|>Tim has 86 shells. Tim gets
  6. chat 12.1
    Answer with the number only.<|system_end|><|user_start|>Sue has 302 apples. Sue gets
  7. worksheet 12.02
    <|assistant_end|><|bos|>Tom has 87274 apples. Tom gets
  8. python 11.95
    :⏎```⏎9⏎3⏎```<|assistant_end|><|bos|># test_items.py⏎from
  9. python 11.89
    ⏎7 plus 6 is 13.<|bos|># test_items.py⏎from
  10. chat 11.85
    Answer with the number only.<|system_end|><|user_start|>Lily has 371 apples. Lily gets
  11. chat 11.81
    Answer with the number only.<|system_end|><|user_start|>Sue has 281 apples. Sue gets
  12. python 11.78
    .py` passes: all tests passed.<|assistant_end|><|bos|># test_stats.py⏎from