Feature #928
Fires on “p” (not labeled yet: read the examples)
- Fires on
- 0.32% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 91%
- story 6%
- chat 1%
- worksheet 1%
- python 0%
Makes the model say
youred plusve thankfuler<|arg|>et
Makes it not say
So( dark fun I very turned three
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
agent 3.4make new friends.<|bos|><|system_start|>tools: edit, glob, grep, question, read
-
agent 3.38with each other.<|bos|><|system_start|>tools: edit, glob, grep, question, read
-
agent 3.37named Ollie.<|assistant_end|><|bos|><|system_start|>tools: edit, glob, grep, question, read
-
agent 3.36day long.<|assistant_end|><|bos|><|system_start|>tools: edit, glob, grep, question, read
-
agent 3.35named Tim.<|assistant_end|><|bos|><|system_start|>tools: edit, glob, grep, question, read
-
agent 3.35297<|assistant_end|><|bos|><|system_start|>tools: edit, glob, grep, question, read
-
agent 3.35pass now.<|assistant_end|><|bos|><|system_start|>tools: edit, glob, grep, question, read
-
agent 3.34the park.<|assistant_end|><|bos|><|system_start|>tools: edit, glob, grep, question, read
-
agent 3.34615<|bos|><|system_start|>tools: edit, glob, grep, question, read
-
agent 3.33976<|assistant_end|><|bos|><|system_start|>tools: edit, glob, grep, question, read
-
agent 3.33-lab.<|assistant_end|><|bos|><|system_start|>tools: edit, glob, grep, question, read
-
agent 3.32from their friends.<|bos|><|system_start|>tools: edit, glob, grep, question, read