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

Feature #2749

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

Fires on
0.31% of tokens
Mostly in
worksheet
Layer
after block 4

Where it fires

  • worksheet 83%
  • chat 9%
  • story 4%
  • agent 3%
  • python 0%

Makes the model say

oseThank thanked1<|tool_call_start|> hugsletter him

Makes it not say

They way Now timesest best Theners

Strongest examples

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

  1. worksheet 10.29
    + 62486 = 1443
  2. worksheet 9.69
    apples and found 38056 more. Now Ben has 3809
  3. worksheet 9.25
    The sum of 61 and 93008 is 9306
  4. worksheet 9.23
    The sum of 29 and 75899 is 7592
  5. worksheet 9.16
    + 15834 = 7853
  6. worksheet 9.11
    add 16 and 34070, you get 3408
  7. worksheet 9.07
    + 18559 = 1859
  8. worksheet 9.04
    books and found 94109 more. Now Tom has 9411
  9. worksheet 9.0
    ⏎The sum of 2 and 69817 is 6981
  10. worksheet 8.88
    ⏎The sum of 3 and 64507 is 6451
  11. worksheet 8.86
    stickers and found 68993 more. Now Anna has 6900
  12. worksheet 8.81
    The sum of 13 and 45864 is 4587