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

Feature #109

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

Fires on
0.61% of tokens
Mostly in
chat
Layer
after block 4

Where it fires

  • chat 41%
  • agent 29%
  • worksheet 13%
  • python 12%
  • story 5%

Makes the model say

0e It game B called l noise

Makes it not say

y the has passesAdded_*.ndered her

Strongest examples

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

  1. agent 3.35
    + b) /
  2. chat 3.35
    + 93<|tool_call_end|>
  3. agent 3.35
    + 1<|tool_end|><|assistant_start|>
  4. chat 3.35
    + 6<|tool_call_end|><|assistant_end|>
  5. chat 3.35
    + 35<|tool_call_end|>
  6. chat 3.35
    + 9?<|user_end|>
  7. python 3.35
    + b⏎⏎$
  8. chat 3.35
    + 920
  9. python 3.35
    + "name" +
  10. chat 3.35
    + 619
  11. chat 3.35
    + 500
  12. agent 3.35
    + name + "!"⏎