Skip to content
mini-lab
Features of prelude-2

Feature #3774

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

Fires on
0.08% of tokens
Mostly in
chat
Layer
after block 4

Where it fires

  • chat 43%
  • story 24%
  • worksheet 22%
  • agent 8%
  • python 2%

Makes the model say

O why all⏎ know moral plus O

Makes it not say

definesaeiou nearbyDo function<|tool_end|>6

Strongest examples

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

  1. chat 1.4
    : calculator<|system_end|><|user_start|>Compute 6+4<|user_end|><|assistant_start|><|tool_call_start|>calculator<|arg|>expression=6 +
  2. chat 1.39
    46 and 5.<|user_end|><|assistant_start|><|tool_call_start|>calculator<|arg|>expression=46 +
  3. chat 1.36
    of 24 and 1.<|user_end|><|assistant_start|><|tool_call_start|>calculator<|arg|>expression=24 +
  4. chat 1.34
    : calculator<|system_end|><|user_start|>Compute 6+4<|user_end|><|assistant_start|><|tool_call_start|>calculator<|arg|>expression=6 +
  5. chat 1.32
    calculator<|system_end|><|user_start|>1 + 5 = ?<|user_end|><|assistant_start|><|tool_call_start|>calculator<|arg|>expression=1 +
  6. chat 1.31
    add 76 and 20.<|user_end|><|assistant_start|><|tool_call_start|>calculator<|arg|>expression=76 +
  7. chat 1.29
    What's 68 + 1?<|user_end|><|assistant_start|><|tool_call_start|>calculator<|arg|>expression=68 +
  8. chat 1.28
    to know 97 + 0.<|user_end|><|assistant_start|><|tool_call_start|>calculator<|arg|>expression=97 +
  9. story 1.28
    Tommy was a curious three year old. One day, he ventured off into a
  10. story 1.27
    Amy. Amy liked to play in the park. One day, she discovered a shiny coin
  11. story 1.27
    the toy all day. The other dog saw this and felt envious too. But then
  12. python 1.26
    Tree"))⏎⏎$ python3 app.py⏎8⏎8⏎tree⏎$