Skip to content
mini-lab
Features of prelude-2

Feature #1700

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

Fires on
0.53% of tokens
Mostly in
agent
Layer
after block 4

Where it fires

  • agent 54%
  • story 28%
  • chat 13%
  • worksheet 4%
  • python 1%

Makes the model say

scared old that fast well what why

Makes it not say

years more counts project⏎⏎ Python"`.ting

Strongest examples

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

  1. story 2.52
    it blew Mittens up into a tall tree. She was scared and cried for help.
  2. chat 2.42
    much is 2 + 96?<|user_end|><|assistant_start|><|tool_call_start|>calculator<|arg|>expression=2 +
  3. chat 2.32
    What's 68 + 1?<|user_end|><|assistant_start|><|tool_call_start|>calculator<|arg|>expression=68 +
  4. chat 2.32
    <|assistant_end|><|user_start|>And add 6 to that?<|user_end|><|assistant_start|><|tool_call_start|>calculator<|arg|>expression=172
  5. chat 2.31
    : calculator<|system_end|><|user_start|>Compute 6+4<|user_end|><|assistant_start|><|tool_call_start|>calculator<|arg|>expression=6 +
  6. chat 2.3
    <|user_start|>Now add 7 to the result.<|user_end|><|assistant_start|><|tool_call_start|>calculator<|arg|>expression=65 +
  7. chat 2.3
    7.<|assistant_end|><|user_start|>Plus 10?<|user_end|><|assistant_start|><|tool_call_start|>calculator<|arg|>expression=77 +
  8. chat 2.3
    <|assistant_end|><|user_start|>And add 0 to that?<|user_end|><|assistant_start|><|tool_call_start|>calculator<|arg|>expression=82 +
  9. story 2.28
    . He feels the water in his mouth and nose. He is scared and sorry. He wants
  10. chat 2.28
    <|user_start|>Now add 1 to the result.<|user_end|><|assistant_start|><|tool_call_start|>calculator<|arg|>expression=49 +
  11. agent 2.27
    <|tool_call_start|>edit<|arg|>path=tools.py<|arg|>oldString=first_letter
  12. chat 2.27
    know 86 + 60.<|user_end|><|assistant_start|><|tool_call_start|>calculator<|arg|>expression=86 +