Feature #1884
Fires on “utils” (not labeled yet: read the examples)
- Fires on
- 0.58% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 85%
- python 8%
- story 4%
- chat 3%
- worksheet 0%
Makes the model say
ne table sunmy.'s danced eat
Makes it not say
out told2 about onlyoldStringdivOUNT
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
agent 3.18the average of a and b.<|user_end|><|assistant_start|><|tool_call_start|>read<|arg|>path=utils.py<|tool_call_end|><|assistant_end|>
-
agent 2.89the smaller of a and b"<|user_end|><|assistant_start|><|tool_call_start|>read<|arg|>path=utils.py<|tool_call_end|><|assistant_end|>
-
agent 2.86the number of characters in s.<|user_end|><|assistant_start|><|tool_call_start|>read<|arg|>path=strings.py<|tool_call_end|><|assistant_end|>
-
agent 2.86doubles every number in nums.<|user_end|><|assistant_start|><|tool_call_start|>read<|arg|>path=utils.py<|tool_call_end|><|assistant_end|>
-
agent 2.85the smaller of a and b.<|user_end|><|assistant_start|><|tool_call_start|>read<|arg|>path=utils.py<|tool_call_end|><|assistant_end|>
-
agent 2.84the first letter of s please"<|user_end|><|assistant_start|><|tool_call_start|>read<|arg|>path=utils.py<|tool_call_end|><|assistant_end|>
-
agent 2.83the sum of the numbers in nums<|user_end|><|assistant_start|><|tool_call_start|>read<|arg|>path=utils.py<|tool_call_end|><|assistant_end|>
-
agent 2.82: it returns a minus b<|user_end|><|assistant_start|><|tool_call_start|>write<|arg|>path=utils.py<|arg|>content
-
agent 2.82It returns the average of nums.<|user_end|><|assistant_start|><|tool_call_start|>read<|arg|>path=lists.py<|tool_call_end|><|assistant_end|>
-
agent 2.78returns every number of nums doubled.<|user_end|><|assistant_start|><|tool_call_start|>read<|arg|>path=utils.py<|tool_call_end|><|assistant_end|>
-
agent 2.77What's in utils.py?<|user_end|><|assistant_start|><|tool_call_start|>read<|arg|>path=utils.py<|tool_call_end|><|assistant_end|>
-
agent 2.75the average of a and b.<|user_end|><|assistant_start|><|tool_call_start|>read<|arg|>path=utils.py<|tool_call_end|><|assistant_end|>