Skip to content
mini-lab
Features of prelude-2

Feature #1981

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

Fires on
0.49% of tokens
Mostly in
agent
Layer
after block 4

Where it fires

  • agent 42%
  • chat 22%
  • story 19%
  • worksheet 15%
  • python 2%

Makes the model say

you happened was happen? is_<|assistant_end|>

Makes it not say

example python])`appedNohttpsbedally

Strongest examples

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

  1. chat 2.71
    <|system_end|><|system_start|>Do not use the calculator.<|system_end|><|user_start|>Could you tell me what 365
  2. chat 2.62
    <|system_end|><|system_start|>Do not use the calculator.<|system_end|><|user_start|>Could you tell me what 365
  3. worksheet 2.59
    Now Tim has 326 stars.⏎Could you tell me what 5 +
  4. worksheet 2.59
    Now Sue has 136 cookies.⏎Could you tell me what 6 +
  5. chat 2.59
    <|bos|><|system_start|>Answer with the number only.<|system_end|><|user_start|>Could you tell me what 453
  6. worksheet 2.58
    26, you get 40.⏎Could you tell me what 52 +
  7. worksheet 2.57
    0 = 9262.⏎Could you tell me what 99 +
  8. worksheet 2.53
    0 = 55310.⏎Could you tell me what 225
  9. worksheet 2.52
    + 6 = 683.⏎Could you tell me what 533
  10. worksheet 2.52
    3 = 37666.⏎Could you tell me what 289
  11. worksheet 2.51
    8 equals 17107.⏎Could you tell me what 594
  12. worksheet 2.51
    only tell short stories and add numbers.<|assistant_end|><|bos|>Could you tell me what 922