Skip to content
mini-lab
Features of prelude-2

Feature #3917

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

Fires on
0.75% of tokens
Mostly in
agent
Layer
after block 4

Where it fires

  • agent 56%
  • chat 20%
  • python 14%
  • story 5%
  • worksheet 4%

Makes the model say

<|arg|> table lookedub near watch functions

Makes it not say

⏎tet":red):<|bos|>

Strongest examples

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

  1. agent 1.63
    .py. It returns the first letter of s.<|user_end|><|assistant_start|><|tool_call_start|>read<|arg|>path=lib
  2. agent 1.57
    stats.py: it returns the average of nums.<|user_end|><|assistant_start|><|tool_call_start|>read<|arg|>path=stats
  3. agent 1.53
    to greetings.py: it makes s upper case.<|user_end|><|assistant_start|><|tool_call_start|>read<|arg|>path=greetings
  4. agent 1.52
    <|user_start|>Please which functions are in strings.py?<|user_end|><|assistant_start|><|tool_call_start|>read<|arg|>path=strings
  5. agent 1.51
    py. It returns the average of a and b.<|user_end|><|assistant_start|><|tool_call_start|>read<|arg|>path=utils
  6. agent 1.51
    , what does negate return?<|user_end|><|assistant_start|><|tool_call_start|>read<|arg|>path=ops
  7. agent 1.51
    <|system_end|><|user_start|>explain product from number_utils.py<|user_end|><|assistant_start|><|tool_call_start|>read<|arg|>path=number
  8. agent 1.5
    <|user_start|>What's in items.py?<|user_end|><|assistant_start|><|tool_call_start|>read<|arg|>path=items
  9. agent 1.49
    py: it returns the sum of a and b.<|user_end|><|assistant_start|><|tool_call_start|>read<|arg|>path=math
  10. chat 1.49
    04 + 58?<|user_end|><|assistant_start|><|tool_call_start|>calculator<|arg|>expression=204 +
  11. story 1.48
    Mia says.⏎Tom shakes his head. He knows the truth. Flow
  12. chat 1.48
    3 and 3 for me?<|user_end|><|assistant_start|><|tool_call_start|>calculator<|arg|>expression=53 + 3