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

Feature #3367

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

Fires on
0.99% of tokens
Mostly in
worksheet
Layer
after block 4

Where it fires

  • worksheet 57%
  • chat 30%
  • agent 10%
  • story 2%
  • python 1%

Makes the model say

? I One function()evens men

Makes it not say

dayient wentk important songMomread

Strongest examples

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

  1. worksheet 4.25
    7+8: 7+8=15, 15791
  2. worksheet 4.07
    ⏎7+8: 7+8=15, 15814
  3. chat 4.02
    <|think_start|>8+9: 8+9=17, 17<|think_end|>The answer
  4. worksheet 4.02
    6+7: 6+7=13, 13810
  5. chat 4.01
    <|think_start|>8+9: 8+9=17, 17<|think_end|>The answer
  6. worksheet 3.95
    ⏎2+8: 2+8=10, 10649
  7. chat 3.93
    <|think_start|>8+5: 8+5=13, 13<|think_end|>13
  8. worksheet 3.91
    ⏎6+8: 6+8=14, 14169
  9. worksheet 3.89
    ⏎7+8: 7+8=15, 15864
  10. chat 3.81
    <|think_start|>2+8: 2+8=10, 10<|think_end|>The answer
  11. worksheet 3.79
    ⏎6+9: 6+9=15, 15686
  12. chat 3.78
    <|think_start|>8+2: 8+2=10, 10<|think_end|>The answer