Skip to content
mini-lab
Features of prelude-2

Feature #3450

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

Fires on
1.02% of tokens
Mostly in
worksheet
Layer
after block 4

Where it fires

  • worksheet 69%
  • chat 20%
  • story 6%
  • agent 3%
  • python 2%

Makes the model say

- + but. fall byS in

Makes it not say

pyer opened knows looked PythonmanyPlease

Strongest examples

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

  1. worksheet 3.55
    8 = 7561.⏎Tim had 94 stars and found 2
  2. worksheet 3.51
    = 53757⏎Anna had 263 flowers and found 1
  3. worksheet 3.42
    67.⏎Lily had 66 apples and found 6
  4. worksheet 3.34
    answer is 109.<|assistant_end|><|bos|>Lily had 70 flowers and found 6
  5. worksheet 3.25
    102579.⏎Mia had 505 flowers and found 9
  6. worksheet 3.23
    , 4]`.⏎<|bos|>Tom had 84299 stars and found 2
  7. worksheet 3.21
    The tests pass now.<|assistant_end|><|bos|>Tim had 5298 stars and found 6
  8. worksheet 3.18
    = 75834.⏎Mia had 35 apples and found 4
  9. worksheet 3.17
    + 0 = 46.⏎Lucy had 54 shells and found 4
  10. worksheet 3.16
    = 1377⏎Ben had 41 apples and found 3
  11. worksheet 3.15
    70 = 77.⏎Tom had 640 shells and found 2
  12. worksheet 3.13
    8 = 97079.⏎Lily had 3 apples and found 8