Feature #314
Fires on “You” (not labeled yet: read the examples)
- Fires on
- 0.61% of tokens
- Mostly in
- chat
- Layer
- after block 4
Where it fires
- chat 47%
- agent 26%
- worksheet 10%
- story 8%
- python 8%
Makes the model say
importresult<|assistant_start|>:/// calculator For<|arg|>
Makes it not say
, withour b and went square co
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 2.86with that. I only know how to tell stories and add numbers.<|assistant_end|><|bos|>Once upon a
-
chat 2.82with that. I only know how to tell stories and add numbers.<|assistant_end|><|user_start|>Summ
-
chat 2.82with that. I only know how to tell stories and add numbers.<|assistant_end|><|user_start|>tell me
-
agent 2.77<|system_end|><|system_start|>You are an AI agent
-
agent 2.77<|system_end|><|system_start|>You are an AI agent
-
agent 2.77<|system_end|><|system_start|>You are an AI agent
-
agent 2.77<|system_end|><|system_start|>You are an AI agent
-
agent 2.77<|system_end|><|system_start|>You are an AI agent
-
agent 2.77<|system_end|><|system_start|>You are an AI agent
-
agent 2.77<|system_end|><|system_start|>You are an AI agent
-
agent 2.77<|system_end|><|system_start|>You are an AI agent
-
agent 2.77<|system_end|><|system_start|>You are an AI agent