Feature #1648
Fires on “ m” (not labeled yet: read the examples)
- Fires on
- 1.21% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 72%
- python 12%
- story 12%
- chat 4%
- worksheet 1%
Makes the model say
ONshpathet!")ye checksar
Makes it not say
to one He on and return his a
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
story 3.06two, three, four, five, six, seven, eight, nine
-
agent 2.86, websearch, write, execute, memory<|system_end|><|system_start|>
-
agent 2.83, websearch, write, execute, memory, cont
-
agent 2.81, websearch, write, execute, memory<|system_end|><|system_start|>
-
agent 2.8, websearch, write, execute, memory, cont
-
python 2.78OUD") == "cloud"⏎assert lower_case("HELLO")
-
story 2.78to protect you."⏎Jack nodded his head and said, "Yes mummy. I will
-
story 2.77to use my machine?" he says.⏎"We are sorry, mister," says Anna
-
agent 2.77, websearch, write, execute, memory, cont
-
agent 2.73, websearch, write, execute, memory<|system_end|><|system_start|>
-
agent 2.73, websearch, write, execute, memory<|system_end|><|system_start|>
-
agent 2.72, websearch, write, execute, memory<|system_end|><|system_start|>