Skip to content
mini-lab
Features of prelude-2

Feature #87

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

Fires on
0.77% of tokens
Mostly in
story
Layer
after block 4

Where it fires

  • story 43%
  • chat 24%
  • agent 19%
  • worksheet 13%
  • python 1%

Makes the model say

ure idea andeven faster lot back getting

Makes it not say

=⏎ Fix theON?"pped?

Strongest examples

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

  1. chat 3.53
    <|assistant_start|><|think_start|>8+9: 8+9=17, 17<|think_end|>The
  2. chat 3.51
    <|assistant_start|><|think_start|>9+1: 9+1=10, 10<|think_end|>The
  3. chat 3.41
    <|assistant_start|><|think_start|>9+8: 9+8=17, 17<|think_end|>The
  4. chat 3.3
    <|assistant_start|><|think_start|>9+9: 9+9=18, 18<|think_end|>The
  5. chat 3.28
    4+7: 4+7+1=12, 121<|think_end|>
  6. chat 3.27
    <|assistant_start|><|think_start|>8+9: 8+9=17, 17<|think_end|>The
  7. chat 3.24
    +8: 8+8=16, 167<|think_end|>
  8. worksheet 3.24
    7+8: 7+8=15, 1579
  9. chat 3.21
    9+8: 9+8+1=18, 183<|think_end|>
  10. worksheet 3.2
    ⏎8+7: 8+7=15, 1592
  11. chat 3.19
    9⏎9+8: 9+8=17, 179<|think_end|>
  12. chat 3.17
    9+7: 9+7+1=17, 1747