Feature #1313
Fires on “ you” (not labeled yet: read the examples)
- Fires on
- 0.26% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 85%
- story 8%
- chat 4%
- python 1%
- worksheet 1%
Makes the model say
ss⏎."? thirstyIfOf but
Makes it not say
home great named otherHiHello more impress
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
story 1.75you can make lots of
-
agent 1.75you?<|user_end|><|assistant_start|>I
-
chat 1.75you tell me what
-
worksheet 1.75you get 12
-
story 1.75you help me?" Mia
-
agent 1.75you run run.py
-
story 1.75you too, Tom.
-
chat 1.75you tell me a story
-
story 1.75you get some blueberries
-
chat 1.75you today?<|assistant_end|><|user_start|>
-
story 1.75you Fluffy. Do you
-
agent 1.7, question, read, shell, skill, subagent, webfetch