Feature #1811
Fires on “appy” (not labeled yet: read the examples)
- Fires on
- 0.18% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 62%
- chat 22%
- story 14%
- worksheet 2%
- python 1%
Makes the model say
time day topy isselfTrue done
Makes it not say
usnums plusiAddut_ +
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 3.46.<|assistant_end|><|bos|><|user_start|>Great, thanks!<|user_end|><|assistant_start|>Happy to help!<|assistant_end|>
-
chat 3.39helpful assistant.<|system_end|><|user_start|>Great, thanks!<|user_end|><|assistant_start|>Happy to help!<|assistant_end|>
-
chat 3.27back to playing normally.<|bos|><|user_start|>thanks<|user_end|><|assistant_start|>Happy to help!<|assistant_end|>
-
chat 3.25py`.<|assistant_end|><|bos|><|user_start|>Great, thanks!<|user_end|><|assistant_start|>Happy to help!<|assistant_end|>
-
chat 3.25files in this project.<|assistant_end|><|bos|><|user_start|>thanks<|user_end|><|assistant_start|>Happy to help!<|assistant_end|>
-
chat 3.24helpful assistant.<|system_end|><|user_start|>Great, thanks!<|user_end|><|assistant_start|>Happy to help!<|assistant_end|>
-
chat 3.23story!<|assistant_end|><|bos|><|user_start|>Great, thanks!<|user_end|><|assistant_start|>Happy to help!<|assistant_end|>
-
chat 3.23laptop by mini-lab.<|assistant_end|><|user_start|>Thanks!<|user_end|><|assistant_start|>Happy to help!<|assistant_end|>
-
chat 3.2<|assistant_end|><|bos|><|system_start|>tools: calculator<|system_end|><|user_start|>thanks<|user_end|><|assistant_start|>Happy to help!<|assistant_end|>
-
chat 3.19py`.<|assistant_end|><|bos|><|user_start|>Great, thanks!<|user_end|><|assistant_start|>Happy to help!<|assistant_end|>
-
chat 3.1743.<|assistant_end|><|user_start|>Great, thanks!<|user_end|><|assistant_start|>Happy to help!<|assistant_end|>
-
chat 3.17answer is 10.<|assistant_end|><|user_start|>thanks<|user_end|><|assistant_start|>Happy to help!<|assistant_end|>