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

Feature #2908

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

Fires on
0.28% of tokens
Mostly in
chat
Layer
after block 4

Where it fires

  • chat 75%
  • story 21%
  • agent 3%
  • python 1%
  • worksheet 0%

Makes the model say

from of stopped they ons out

Makes it not say

de])Heyfereadouenv or

Strongest examples

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

  1. story 3.05
    big and red. The train was happy to help.⏎On the way, the train saw
  2. story 2.89
    , "Yes, Tim. That pie looks yummy!"⏎On the way home, Tim's
  3. chat 2.67
    0+7=7, 766<|think_end|>The answer is 76
  4. story 2.65
    ball that her mom said was perfect for the trip.⏎On the way, they stopped at
  5. chat 2.64
    6+9=15, 15<|think_end|>The answer is 15
  6. chat 2.63
    4+1=14, 1430<|think_end|>The answer is 14
  7. chat 2.56
    0+1=10, 1006<|think_end|>The answer is 10
  8. chat 2.55
    27+758=<|user_end|><|assistant_start|>The answer is 15
  9. story 2.54
    Mia was very happy and put on her yellow coat.⏎On the way to grandma's house
  10. chat 2.54
    9+8=17, 172<|think_end|>The answer is 17
  11. chat 2.54
    +0=1, 131<|think_end|>The answer is 13
  12. chat 2.52
    +1=2, 210<|think_end|>The answer is 21