Skip to content
mini-lab
Features of prelude-2

Feature #371

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

Fires on
0.34% of tokens
Mostly in
agent
Layer
after block 4

Where it fires

  • agent 93%
  • story 3%
  • chat 2%
  • python 1%
  • worksheet 0%

Makes the model say

accant too himutils a anyHi

Makes it not say

of stop If named stopped shouldIf towards

Strongest examples

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

  1. agent 3.12
    =python3 test_calc.py<|tool_call_end|><|assistant_end|><|tool_start|>Traceback (most recent call last):
  2. agent 3.12
    =python3 test_calc.py<|tool_call_end|><|assistant_end|><|tool_start|>Traceback (most recent call last):
  3. agent 3.12
    python3 test_tools.py<|tool_call_end|><|assistant_end|><|tool_start|>Traceback (most recent call last):
  4. agent 3.09
    test_text.py<|tool_call_end|><|assistant_end|><|tool_start|>Traceback (most recent call last):
  5. agent 3.03
    py<|tool_call_end|><|assistant_end|><|tool_start|>Traceback (most recent call last):
  6. agent 2.94
    _strings.py<|tool_call_end|><|assistant_end|><|tool_start|>Traceback (most recent call last):
  7. agent 2.81
    command=python3 test_tools.py<|tool_call_end|><|assistant_end|><|tool_start|>Traceback (most recent call last):
  8. agent 2.79
    <|tool_call_end|><|assistant_end|><|tool_start|>Traceback (most recent call last):
  9. agent 2.79
    <|tool_call_end|><|assistant_end|><|tool_start|>Traceback (most recent call last):
  10. agent 2.79
    command=python3 test_tools.py<|tool_call_end|><|assistant_end|><|tool_start|>Traceback (most recent call last):
  11. agent 2.78
    .py<|tool_call_end|><|assistant_end|><|tool_start|>Traceback (most recent call last):
  12. agent 2.78
    .py<|tool_call_end|><|assistant_end|><|tool_start|>Traceback (most recent call last):