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

Feature #2661

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

Fires on
1.11% of tokens
Mostly in
worksheet
Layer
after block 4

Where it fires

  • worksheet 67%
  • chat 25%
  • python 4%
  • agent 3%
  • story 1%

Makes the model say

<|arg|> absoluteful0if Maxname",

Makes it not say

$ lot everywhere home anythinglire there

Strongest examples

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

  1. chat 4.44
    2⏎37+03: 7+3=10, 0
  2. worksheet 4.39
    12⏎67+73: 7+3=10, 0
  3. chat 4.28
    <|think_start|>007+973: 7+3=10, 0
  4. worksheet 4.26
    277+7043: 7+3=10, 0
  5. chat 4.21
    3.<|user_end|><|assistant_start|><|think_start|>7+3: 7+3=10, 1
  6. worksheet 4.18
    ⏎007+443: 7+3=10, 0
  7. chat 4.05
    4⏎63+57: 3+7=10, 0
  8. worksheet 4.05
    7: 3+7=10, 0
  9. worksheet 4.05
    7: 3+7=10, 0
  10. chat 4.02
    <|user_end|><|assistant_start|><|think_start|>37+23: 7+3=10, 0
  11. worksheet 4.0
    ⏎47+72: 7+2+1=10, 0
  12. worksheet 3.98
    457+8493: 7+3=10, 0