Feature #1540
Fires on “p” (not labeled yet: read the examples)
- Fires on
- 0.65% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 74%
- story 21%
- chat 3%
- worksheet 1%
- python 0%
Makes the model say
<|arg|> SopatternSoitename[:: goodbye
Makes it not say
b t6 c7 It's is
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
agent 3.9the name of max_of to larger."<|user_end|><|assistant_start|><|tool_call_start|>grep<|arg|>pattern=max
-
agent 3.83Use Eva for USER.<|user_end|><|assistant_start|><|tool_call_start|>grep<|arg|>pattern=U
-
agent 3.77"Where can I find is_divisible?"<|user_end|><|assistant_start|><|tool_call_start|>grep<|arg|>pattern=def
-
agent 3.77<|system_end|><|user_start|>"Where can I find double?"<|user_end|><|assistant_start|><|tool_call_start|>grep<|arg|>pattern=def
-
agent 3.76.py that returns the average of a and b.<|user_end|><|assistant_start|><|tool_call_start|>read<|arg|>path=main
-
agent 3.74"Explain the function how_many."<|user_end|><|assistant_start|><|tool_call_start|>grep<|arg|>pattern=def
-
agent 3.73<|user_start|>"how does is_even work?"<|user_end|><|assistant_start|><|tool_call_start|>grep<|arg|>pattern=def
-
agent 3.73_v2 instead of last_letter.<|user_end|><|assistant_start|><|tool_call_start|>grep<|arg|>pattern=last
-
agent 3.68_item to my_max_item."<|user_end|><|assistant_start|><|tool_call_start|>grep<|arg|>pattern=max
-
agent 3.68.<|system_end|><|user_start|>"what does loud do?"<|user_end|><|assistant_start|><|tool_call_start|>grep<|arg|>pattern=def
-
agent 3.68.<|system_end|><|user_start|>find the function magnitude.<|user_end|><|assistant_start|><|tool_call_start|>grep<|arg|>pattern=def
-
agent 3.66.<|system_end|><|user_start|>"where is combine defined?"<|user_end|><|assistant_start|><|tool_call_start|>grep<|arg|>pattern=def