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

Feature #1610

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

Fires on
0.17% of tokens
Mostly in
agent
Layer
after block 4

Where it fires

  • agent 30%
  • worksheet 30%
  • chat 25%
  • story 10%
  • python 4%

Makes the model say

from+ that. together From result has

Makes it not say

"!"rngmin}ings returns]

Strongest examples

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

  1. chat 4.1
    <|user_end|><|assistant_start|><|think_start|>16+77: 6+7=13, 3
  2. worksheet 4.0
    ⏎087+536: 7+6=13, 3
  3. chat 3.89
    <|user_end|><|assistant_start|><|think_start|>17+76: 7+6=13, 3
  4. chat 3.88
    <|user_end|><|assistant_start|><|think_start|>87+66: 7+6=13, 3
  5. chat 3.83
    9, 9⏎6+7: 6+7=13, 1
  6. worksheet 3.82
    ⏎626+007: 6+7=13, 3
  7. chat 3.81
    <|think_start|>307+026: 7+6=13, 3
  8. worksheet 3.8
    716+0097: 6+7=13, 3
  9. chat 3.8
    <|user_end|><|assistant_start|><|think_start|>46+67: 6+7=13, 3
  10. worksheet 3.77
    56+00007: 6+7=13, 3
  11. chat 3.77
    <|think_start|>446+047: 6+7=13, 3
  12. worksheet 3.76
    is?⏎06+77: 6+7=13, 3