Skip to content
mini-lab
Features of prelude-2

Feature #214

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

Fires on
0.46% of tokens
Mostly in
agent
Layer
after block 4

Where it fires

  • agent 78%
  • story 14%
  • chat 4%
  • python 2%
  • worksheet 2%

Makes the model say

again at her - wouldan say herself

Makes it not say

Iassert⏎()Hi The they They

Strongest examples

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

  1. agent 2.34
    , what does negate return?<|user_end|><|assistant_start|><|tool_call_start|>read<|arg|>path=ops.py
  2. agent 2.31
    , 0, 3]))<|tool_end|><|assistant_start|><|tool_call_start|>edit<|arg|>path=main.py
  3. agent 2.3
    ⏎2: return a + b<|tool_end|><|assistant_start|><|tool_call_start|>edit<|arg|>path=core.py
  4. agent 2.29
    0: return x + 1<|tool_end|><|assistant_start|><|tool_call_start|>edit<|arg|>path=number_utils
  5. agent 2.28
    9]) == 15⏎<|tool_end|><|assistant_start|><|tool_call_start|>edit<|arg|>path=lists.py
  6. agent 2.25
    3, 1, 2]))<|tool_end|><|assistant_start|><|tool_call_start|>edit<|arg|>path=hello.py
  7. agent 2.23
    , 8, 2]))<|tool_end|><|assistant_start|><|tool_call_start|>edit<|arg|>path=script.py
  8. agent 2.21
    ⏎2: return s + s<|tool_end|><|assistant_start|><|tool_call_start|>edit<|arg|>path=text.py
  9. agent 2.2
    .py, what does positive return?<|user_end|><|assistant_start|><|tool_call_start|>read<|arg|>path=run.py
  10. agent 2.2
    10: return a - b<|tool_end|><|assistant_start|><|tool_call_start|>edit<|arg|>path=core.py
  11. agent 2.18
    + 2 for n in nums]<|tool_end|><|assistant_start|><|tool_call_start|>edit<|arg|>path=lists.py
  12. agent 2.18
    9, 2, 3]))<|tool_end|><|assistant_start|><|tool_call_start|>edit<|arg|>path=hello.py