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

Feature #2423

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

Fires on
0.51% of tokens
Mostly in
worksheet
Layer
after block 4

Where it fires

  • worksheet 31%
  • agent 30%
  • python 18%
  • chat 15%
  • story 6%

Makes the model say

letters time WeItainightvowels

Makes it not say

>>> upon range# exampleroom there sunny

Strongest examples

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

  1. chat 5.88
    <|user_end|><|assistant_start|><|think_start|>16+77: 6+7=13, 3
  2. chat 5.79
    9, 9⏎6+7: 6+7=13, 1
  3. chat 5.61
    <|user_end|><|assistant_start|><|think_start|>46+67: 6+7=13, 3
  4. chat 5.6
    <|think_start|>446+047: 6+7=13, 3
  5. worksheet 5.56
    380⏎6+7: 6+7=13, 1
  6. chat 5.55
    <|think_start|>376+007: 6+7=13, 3
  7. chat 5.52
    , 68⏎7+6: 7+6=13, 1
  8. chat 5.5
    <|think_start|>096+667: 6+7=13, 3
  9. chat 5.5
    <|user_end|><|assistant_start|><|think_start|>87+66: 7+6=13, 3
  10. chat 5.48
    7.<|user_end|><|assistant_start|><|think_start|>6+7: 6+7=13, 1
  11. chat 5.48
    6, 6⏎7+6: 7+6=13, 1
  12. worksheet 5.47
    ⏎626+007: 6+7=13, 3