Feature #1562
Fires on โ,โ (not labeled yet: read the examples)
- Fires on
- 0.62% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 79%
- chat 19%
- story 2%
- worksheet 0%
- python 0%
Makes the model say
list files trained Python* passes:Rename
Makes it not say
0ig longr Theyad57
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
agent 1.96: edit, glob, grep, question, read, shell, skill, sub
-
chat 1.92<|assistant_end|><|user_start|>Hi!<|user_end|><|assistant_start|>Hi there! Would you like a story, or some numbers
-
agent 1.89: edit, glob, grep, question, read, shell, skill, sub
-
agent 1.88: edit, glob, grep, question, read, shell, skill, sub
-
agent 1.86: edit, glob, grep, question, read, shell, skill, sub
-
agent 1.85: edit, glob, grep, question, read, shell, skill, sub
-
chat 1.84<|user_start|>hey there<|user_end|><|assistant_start|>Hi there! Would you like a story, or some numbers
-
story 1.82in any trouble."โBut as he jumped, he heard a loud voice from above.
-
agent 1.82: edit, glob, grep, question, read, shell, skill, sub
-
agent 1.81: edit, glob, grep, question, read, shell, skill, sub
-
agent 1.8: edit, glob, grep, question, read, shell, skill, sub
-
agent 1.79: edit, glob, grep, question, read, shell, skill, sub