Feature #1992
Fires on “script” (not labeled yet: read the examples)
- Fires on
- 1.28% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 90%
- story 4%
- worksheet 3%
- chat 2%
- python 1%
Makes the model say
knewstats - soundE wouldutils were
Makes it not say
Hi outcontent soo# back outside
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
agent 3.992, 3]))<|tool_end|><|assistant_start|><|tool_call_start|>edit<|arg|>path=script.py<|arg|>oldString
-
agent 3.95double_all(VALUES))<|tool_end|><|assistant_start|><|tool_call_start|>edit<|arg|>path=script.py<|arg|>oldString
-
agent 3.91def square(x):⏎<|tool_end|><|assistant_start|><|tool_call_start|>edit<|arg|>path=script.py<|arg|>oldString
-
agent 3.898, 2]))<|tool_end|><|assistant_start|><|tool_call_start|>edit<|arg|>path=script.py<|arg|>oldString
-
agent 3.82def factorial(n):⏎<|tool_end|><|assistant_start|><|tool_call_start|>edit<|arg|>path=script.py<|arg|>oldString
-
agent 3.82is_even(SIZE))<|tool_end|><|assistant_start|><|tool_call_start|>edit<|arg|>path=script.py<|arg|>oldString
-
agent 3.81return min(nums)<|tool_end|><|assistant_start|><|tool_call_start|>edit<|arg|>path=lib.py<|arg|>oldString
-
agent 3.8print(positive(SIZE))<|tool_end|><|assistant_start|><|tool_call_start|>edit<|arg|>path=demo.py<|arg|>oldString
-
agent 3.79letter(TEXT))<|tool_end|><|assistant_start|><|tool_call_start|>edit<|arg|>path=script.py<|arg|>oldString
-
agent 3.762, 10))<|tool_end|><|assistant_start|><|tool_call_start|>edit<|arg|>path=hello.py<|arg|>oldString
-
agent 3.75return max(nums)<|tool_end|><|assistant_start|><|tool_call_start|>edit<|arg|>path=utils.py<|arg|>oldString
-
agent 3.74half(10))<|tool_end|><|assistant_start|><|tool_call_start|>edit<|arg|>path=script.py<|arg|>oldString