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

Feature #1588

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

Fires on
0.42% of tokens
Mostly in
worksheet
Layer
after block 4

Where it fires

  • worksheet 72%
  • chat 16%
  • agent 6%
  • story 4%
  • python 2%

Makes the model say

yourrss t cr big sky a

Makes it not say

returnHelloname all listhelpers family surprise

Strongest examples

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

  1. worksheet 7.32
    + 62486 = 1443
  2. worksheet 6.55
    + 18559 = 1859
  3. worksheet 6.47
    + 15834 = 7853
  4. worksheet 6.24
    apples and found 38056 more. Now Ben has 3809
  5. worksheet 6.07
    The sum of 97 and 27853 is 2795
  6. worksheet 6.07
    3 + 39347 = 9743
  7. worksheet 6.05
    apples and found 67867 more. Now Sue has 6793
  8. worksheet 6.0
    The sum of 89 and 67985 is 6807
  9. worksheet 5.97
    books and found 32930 more. Now Sam has 3299
  10. worksheet 5.97
    add 35 and 39915, you get 3995
  11. worksheet 5.89
    406 and 49137, you get 4954
  12. worksheet 5.85
    py`.<|assistant_end|><|bos|>57 plus 59600 is 5965