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

Feature #3378

Fires on “6” (not labeled yet: read the examples)

Fires on
1.12% of tokens
Mostly in
worksheet
Layer
after block 4

Where it fires

  • worksheet 60%
  • chat 22%
  • agent 14%
  • python 3%
  • story 1%

Makes the model say

<|assistant_end|>. played beak boy."positive play

Makes it not say

forever them hereFixanmd1 "

Strongest examples

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

  1. worksheet 3.88
    6+00005: 6+5=1
  2. worksheet 3.74
    +0001: 1+1=2
  3. worksheet 3.71
    +81658: 6+8=1
  4. worksheet 3.65
    4+0006: 4+6+1
  5. worksheet 3.64
    8+0169: 8+9=1
  6. worksheet 3.38
    77705+00772: 5+2=7
  7. worksheet 3.35
    ⏎00005+87094: 5+4=9
  8. worksheet 3.34
    ⏎00006+21049: 6+9=1
  9. worksheet 3.33
    ⏎00006+82075: 6+5=1
  10. worksheet 3.33
    0+0003: 0+3=3
  11. worksheet 3.32
    ⏎00086+72522: 6+2=8
  12. worksheet 3.29
    ⏎00006+58681: 6+1=7