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

Feature #4086

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

Fires on
4.66% of tokens
Mostly in
worksheet
Layer
after block 4

Where it fires

  • worksheet 60%
  • chat 21%
  • agent 9%
  • python 7%
  • story 3%

Makes the model say

0(" toples1 tall8's

Makes it not say

from."Sorry")Thank me item hold

Strongest examples

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

  1. worksheet 4.03
    found 5401 more. Now Sue has 14113 toys.
  2. worksheet 3.8
    4 stars and found 232 more. Now Tim has 326 stars.
  3. worksheet 3.77
    and found 23233 more. Now Tim has 25415
  4. worksheet 3.75
    ⏎So 3592 + 84 = 3676.⏎
  5. python 3.73
    6) == 6⏎assert previous(12) == 11⏎print
  6. worksheet 3.72
    68135 and 9227 is 77362
  7. worksheet 3.72
    add 9238 and 211, you get 9449
  8. worksheet 3.71
    <|assistant_end|><|bos|>511 + 4982 = 5493⏎1
  9. python 3.7
    square, mean⏎⏎assert sub_one(12) == 11⏎assert
  10. worksheet 3.67
    sum of 1332 and 3206 is 4538
  11. worksheet 3.59
    6314.⏎382 plus 3 equals 385.⏎
  12. worksheet 3.57
    9.⏎502 plus 741 equals 1243.<|bos|>