Feature #2888
Fires on “<|bos|>” (not labeled yet: read the examples)
- Fires on
- 0.91% of tokens
- Mostly in
- chat
- Layer
- after block 4
Where it fires
- chat 29%
- story 26%
- agent 24%
- worksheet 11%
- python 10%
Makes the model say
<|tool_call_end|><|user_start|> definednumsOnce} "command
Makes it not say
and This own again too evenThank both
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
story 10.56ways to remember important things, so we don't let others down.<|bos|>Once upon a time
-
agent 10.4careful when sailing and to not go out in bad weather.<|bos|><|system_start|>tools: edit
-
chat 10.39another, there will be fewer disagreements.<|bos|><|system_start|>tools: calculator
-
story 10.39, edit and fix the Python files in this project.<|assistant_end|><|bos|>Once upon a time
-
agent 10.367, 17<|think_end|>The answer is 17.<|assistant_end|><|bos|><|system_start|>tools: edit
-
agent 10.367, 78<|think_end|>The answer is 78.<|assistant_end|><|bos|><|system_start|>tools: edit
-
story 10.34make everyone happy, even when things don't go as planned.<|bos|>Once upon a time
-
story 10.336, 67<|think_end|>The answer is 67.<|assistant_end|><|bos|>One day, a
-
chat 10.32172<|think_end|>The answer is 172.<|assistant_end|><|bos|><|system_start|>tools: calculator
-
chat 10.32that. I only know how to tell stories and add numbers.<|assistant_end|><|bos|><|user_start|>Who made you
-
chat 10.31, create, edit and fix the Python files in this project.<|assistant_end|><|bos|><|user_start|>Hey<|user_end|><|assistant_start|>
-
python 10.31, and he used his strength to help others.<|bos|>`mean_list