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

Feature #914

Fires on “3” (not labeled yet: read the examples)

Fires on
0.40% of tokens
Mostly in
agent
Layer
after block 4

Where it fires

  • agent 53%
  • worksheet 30%
  • chat 10%
  • python 4%
  • story 3%

Makes the model say

One0 return⏎ate+ upGood

Makes it not say

ItTrueFalse typo python back filesup

Strongest examples

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

  1. worksheet 4.0
    1382<|bos|>Once
  2. worksheet 3.98
    5310⏎5+
  3. python 3.98
    3)⏎-3
  4. chat 3.98
    325<|assistant_end|><|bos|>
  5. worksheet 3.98
    312:
  6. worksheet 3.98
    3 + 39
  7. worksheet 3.98
    38 + 9
  8. worksheet 3.98
    3+181
  9. worksheet 3.98
    347:
  10. chat 3.98
    35.<|user_end|><|assistant_start|>
  11. worksheet 3.98
    3515⏎
  12. agent 3.98
    3, 7,