Feature #3793
Fires on “Good” (not labeled yet: read the examples)
- Fires on
- 0.43% of tokens
- Mostly in
- worksheet
- Layer
- after block 4
Where it fires
- worksheet 63%
- agent 31%
- chat 4%
- story 2%
- python 1%
Makes the model say
timesreplaceAll homex Instead isis was
Makes it not say
Sam python your put toOnefromTrue
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 1.7became good friends.<|bos|><|system_start|>tools: calculator<|system_end|><|user_start|>Bye<|user_end|><|assistant_start|>Goodbye! Have
-
story 1.7, there was a charming festival. The festival had many fun things to do. There were games
-
chat 1.63.<|assistant_end|><|bos|><|system_start|>tools: calculator<|system_end|><|user_start|>Good night!<|user_end|><|assistant_start|>Goodbye! Have
-
chat 1.63.<|assistant_end|><|bos|><|system_start|>tools: calculator<|system_end|><|user_start|>Good night!<|user_end|><|assistant_start|>Goodbye! Have
-
worksheet 1.6272988⏎So 824 + 72164 =
-
story 1.61to pretend they were grown-ups and use the things they saw on the shelf.
-
chat 1.591368.<|assistant_end|><|user_start|>See you later!<|user_end|><|assistant_start|>Goodbye! Have
-
chat 1.59<|bos|><|system_start|>tools: calculator<|system_end|><|user_start|>See you later!<|user_end|><|assistant_start|>Goodbye! Have
-
chat 1.57<|bos|><|system_start|>tools: calculator<|system_end|><|user_start|>See you later!<|user_end|><|assistant_start|>Goodbye! Have
-
chat 1.56numbers. Try asking me for a story!<|assistant_end|><|user_start|>Bye<|user_end|><|assistant_start|>Goodbye! Have
-
chat 1.55s always had gentle results.<|bos|><|user_start|>Good night!<|user_end|><|assistant_start|>Goodbye! Have
-
chat 1.54`.⏎<|bos|><|system_start|>tools: calculator<|system_end|><|user_start|>Good night!<|user_end|><|assistant_start|>Goodbye! Have