Feature #1386
Fires on “w” (not labeled yet: read the examples)
- Fires on
- 0.70% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 73%
- story 14%
- chat 7%
- python 5%
- worksheet 1%
Makes the model say
elseusdrite<|arg|>ch idea better
Makes it not say
798 intoultipl theirdef goodbye
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
agent 4.08a function echo that repeats s n times.<|user_end|><|assistant_start|><|tool_call_start|>write<|arg|>path=
-
agent 3.99.py that returns the average of a and b.<|user_end|><|assistant_start|><|tool_call_start|>write<|arg|>path=
-
agent 3.96function last_letter that returns the last letter of s<|user_end|><|assistant_start|><|tool_call_start|>write<|arg|>path=
-
agent 3.85a function upper_case that makes s upper case.<|user_end|><|assistant_start|><|tool_call_start|>write<|arg|>path=
-
agent 3.79a function smallest that returns the smallest number in nums.<|user_end|><|assistant_start|><|tool_call_start|>write<|arg|>path=
-
story 3.79. The man said he had a big bag of sour candy. He offered Tom some
-
agent 3.76multiply: it returns the product of a and b.<|user_end|><|assistant_start|><|tool_call_start|>write<|arg|>path=
-
agent 3.76file hi.py which prints Let's code"<|user_end|><|assistant_start|><|tool_call_start|>write<|arg|>path=
-
agent 3.75and average that returns the average of a and b.<|user_end|><|assistant_start|><|tool_call_start|>write<|arg|>path=
-
agent 3.72py that prints "Good morning" and run it please<|user_end|><|assistant_start|><|tool_call_start|>write<|arg|>path=
-
agent 3.71) in tools.py that returns the first item.<|user_end|><|assistant_start|><|tool_call_start|>write<|arg|>path=
-
agent 3.68py that prints "Have a nice day" please<|user_end|><|assistant_start|><|tool_call_start|>write<|arg|>path=