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

Feature #3324

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 35%
  • chat 31%
  • python 19%
  • agent 12%
  • story 3%

Makes the model say

09([1ause H [ long

Makes it not say

wish?"])) cutei`:newString]

Strongest examples

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

  1. chat 3.37
    <|think_start|>277+053: 7+3=10, 0
  2. chat 3.36
    <|user_end|><|assistant_start|><|think_start|>37+23: 7+3=10, 0
  3. chat 3.35
    3.<|user_end|><|assistant_start|><|think_start|>7+3: 7+3=10, 1
  4. chat 3.35
    <|think_start|>007+973: 7+3=10, 0
  5. worksheet 3.33
    12⏎67+73: 7+3=10, 0
  6. chat 3.33
    <|think_start|>637+103: 7+3=10, 0
  7. chat 3.24
    2⏎37+03: 7+3=10, 0
  8. worksheet 3.22
    277+7043: 7+3=10, 0
  9. worksheet 3.22
    ⏎007+443: 7+3=10, 0
  10. worksheet 3.2
    ⏎007+283: 7+3=10, 0
  11. worksheet 3.16
    457+8493: 7+3=10, 0
  12. chat 3.13
    <|think_start|>007+953: 7+3=10, 0