Feature #195
Fires on ānā (not labeled yet: read the examples)
- Fires on
- 0.36% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 46%
- python 30%
- story 12%
- chat 11%
- worksheet 1%
Makes the model say
itemspatternite namesight welcome_*.
Makes it not say
Sorry be1 was notest Andape
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 2.77at the festival.<|bos|><|user_start|>Can you book a table for two tonight?<|user_end|><|assistant_start|>
-
story 2.76and saw her cat, Max, playing with the cloth. Max had unwrapped the
-
chat 2.6752.<|assistant_end|><|bos|><|user_start|>Can you book a table for two tonight?<|user_end|><|assistant_start|>
-
chat 2.53<|system_start|>tools: calculator<|system_end|><|user_start|>Can you book a table for two tonight?<|user_end|><|assistant_start|>
-
chat 2.53and add numbers.<|assistant_end|><|user_start|>Can you book a table for two tonight?<|user_end|><|assistant_start|>
-
chat 2.45another story soon.<|assistant_end|><|user_start|>Can you book a table for two tonight?<|user_end|><|assistant_start|>
-
chat 2.43pass now.<|assistant_end|><|bos|><|user_start|>Can you book a table for two tonight?<|user_end|><|assistant_start|>
-
chat 2.42and add numbers.<|assistant_end|><|user_start|>Can you book a table for two tonight?<|user_end|><|assistant_start|>
-
chat 2.39add numbers.<|assistant_end|><|bos|><|user_start|>Can you book a table for two tonight?<|user_end|><|assistant_start|>
-
chat 2.38you today?<|assistant_end|><|user_start|>can you book a table for two tonight<|user_end|><|assistant_start|>Sorry
-
story 2.36would help her to save a very special person.āThe girl was unsure what to
-
chat 2.26another story soon.<|assistant_end|><|user_start|>Can you book a table for two tonight?<|user_end|><|assistant_start|>