Feature #3325
Fires on “<|system_start|>” (not labeled yet: read the examples)
- Fires on
- 0.75% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 48%
- chat 42%
- story 7%
- python 3%
- worksheet 1%
Makes the model say
santp tellsedlathes
Makes it not say
ball because This, one my andThis
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 3.56<|bos|><|system_start|>Answer in one short
-
agent 3.56<|bos|><|system_start|>tools: edit,
-
chat 3.56<|bos|><|system_start|>tools: calculator<|system_end|>
-
chat 3.56<|bos|><|system_start|>tools: calculator<|system_end|>
-
agent 3.56<|bos|><|system_start|>tools: edit,
-
chat 3.56<|bos|><|system_start|>Answer with the number
-
chat 3.56<|bos|><|system_start|>Answer with the number
-
agent 3.37<|assistant_end|><|bos|><|system_start|>tools: edit,
-
agent 3.37<|assistant_end|><|bos|><|system_start|>tools: edit,
-
agent 3.37<|assistant_end|><|bos|><|system_start|>tools: edit,
-
agent 3.37<|assistant_end|><|bos|><|system_start|>tools: edit,
-
agent 3.37<|assistant_end|><|bos|><|system_start|>tools: edit,