Feature #2912
Fires on “<|system_start|>” (not labeled yet: read the examples)
- Fires on
- 1.22% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 55%
- chat 41%
- story 2%
- python 1%
- worksheet 1%
Makes the model say
achightdy makesakingined wantOne
Makes it not say
. the$ named lived them_,
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 2.92a story, or some numbers to add?<|assistant_end|><|bos|><|system_start|>tools: calculator<|system_end|>
-
agent 2.87how to tell stories and add numbers.<|assistant_end|><|bos|><|system_start|>tools: edit,
-
chat 2.87how to tell stories and add numbers.<|assistant_end|><|bos|><|system_start|>tools: calculator<|system_end|>
-
chat 2.84ould you like a story, or some numbers to add?<|assistant_end|><|bos|><|system_start|>Answer with the number
-
chat 2.74ant capabilities<|assistant_end|><|bos|><|system_start|>tools: calculator<|system_end|>
-
agent 2.71325<|assistant_end|><|bos|><|system_start|>tools: edit,
-
chat 2.71story is to always forgive others and learn from our mistakes.<|bos|><|system_start|>tools: calculator<|system_end|>
-
agent 2.68moral of the story is to share and be kind to others.<|bos|><|system_start|>tools: edit,
-
agent 2.67how to add numbers. Try asking me for a story!<|assistant_end|><|bos|><|system_start|>tools: edit,
-
chat 2.66I want to have a marriage when I grow up."<|bos|><|system_start|>tools: calculator<|system_end|>
-
agent 2.65I can only tell short stories and add numbers.<|assistant_end|><|bos|><|system_start|>tools: edit,
-
agent 2.62your parents. They have your best interests at heart.<|bos|><|system_start|>tools: edit,