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

Feature #713

Fires on “1” (not labeled yet: read the examples)

Fires on
0.45% of tokens
Mostly in
worksheet
Layer
after block 4

Where it fires

  • worksheet 63%
  • chat 24%
  • agent 5%
  • story 5%
  • python 2%

Makes the model say

safe<|user_start|> this shadenewowedfromig

Makes it not say

ose nextverFix[::- rockut believe

Strongest examples

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

  1. worksheet 6.32
    8: 4+8+1=13, 3
  2. worksheet 6.2
    2⏎4+8: 4+8+1=13, 1
  3. worksheet 6.01
    4: 9+4=13, 3
  4. worksheet 5.95
    62⏎8+4: 8+4+1=13, 1
  5. worksheet 5.9
    +8: 4+8=12, 1
  6. worksheet 5.8
    74+00039: 4+9=13, 3
  7. worksheet 5.8
    08+244: 8+4+1=13, 3
  8. worksheet 5.79
    50⏎8+4: 8+4+1=13, 1
  9. worksheet 5.77
    244: 9+4=13, 3
  10. chat 5.72
    <|think_start|>079+504: 9+4=13, 3
  11. worksheet 5.69
    98+854: 8+4+1=13, 3
  12. worksheet 5.57
    ⏎54+68: 4+8+1=13, 3