Skip to content
mini-lab
Features of prelude-2

Feature #26

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

Fires on
0.87% of tokens
Mostly in
chat
Layer
after block 4

Where it fires

  • chat 36%
  • story 35%
  • worksheet 14%
  • agent 10%
  • python 5%

Makes the model say

wasSure`,)<|assistant_start|> is<|tool_call_start|>ven

Makes it not say

more little sumt back who rest cream

Strongest examples

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

  1. chat 3.69
    tests pass now.<|assistant_end|><|bos|><|user_start|>Lily has 71 coins. Lily gets 8 more
  2. worksheet 3.61
    6490.<|bos|>Mia has 208 coins. Mia gets 97
  3. worksheet 3.51
    ave a nice day.<|assistant_end|><|bos|>Tim has 86 shells. Tim gets 72
  4. worksheet 3.47
    hi.py`.<|assistant_end|><|bos|>Anna has 88 balls. Anna gets 4 more
  5. chat 3.43
    372.<|assistant_end|><|bos|><|user_start|>Lily has 71 coins. Lily gets 8 more
  6. worksheet 3.4
    18591⏎Tom has 13 cookies. Tom gets 33
  7. chat 3.4
    the number only.<|system_end|><|user_start|>Lily has 371 apples. Lily gets 31
  8. chat 3.38
    is 28.<|assistant_end|><|user_start|>Lily has 47 books. Lily gets 45
  9. worksheet 3.37
    888.⏎Anna has 6195 toys. Anna gets 2 more
  10. chat 3.33
    219.<|assistant_end|><|user_start|>Mia has 29 stickers. Mia gets 54
  11. worksheet 3.32
    5100.⏎Mia has 823 apples. Mia gets 60
  12. chat 3.27
    use the calculator.<|system_end|><|user_start|>Mia has 917 stickers. Mia gets 0 more