Feature #278
Fires on “ at” (not labeled yet: read the examples)
- Fires on
- 0.10% of tokens
- Mostly in
- story
- Layer
- after block 4
Where it fires
- story 40%
- agent 28%
- chat 24%
- python 7%
- worksheet 2%
Makes the model say
No done like amp fix thanis
Makes it not say
<|assistant_end|> lived "$<|think_end|> return<|tool_end|>)`<|system_end|>
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
story 3.24was safe.⏎Tom and Jerry learned that being smart is more important than being fat or thin
-
chat 3.04today?<|assistant_end|><|bos|><|system_start|>tools: calculator<|system_end|><|user_start|>What are you good at?<|user_end|><|assistant_start|>I
-
chat 3.02you<|user_end|><|assistant_start|>Happy to help!<|assistant_end|><|user_start|>what are you good at<|user_end|><|assistant_start|>I'm
-
chat 3.0special that she had earned.<|bos|><|user_start|>What are you good at?<|user_end|><|assistant_start|>I
-
chat 2.99only tell short stories and add numbers.<|assistant_end|><|user_start|>What are you good at?<|user_end|><|assistant_start|>I
-
chat 2.99. Try asking me for a story!<|assistant_end|><|user_start|>What are you good at?<|user_end|><|assistant_start|>I
-
chat 2.97only tell short stories and add numbers.<|assistant_end|><|user_start|>What are you good at?<|user_end|><|assistant_start|>I
-
chat 2.97app.py⏎7⏎6⏎<|bos|><|user_start|>What are you good at?<|user_end|><|assistant_start|>I
-
chat 2.96The answer is 17.<|assistant_end|><|bos|><|user_start|>What are you good at?<|user_end|><|assistant_start|>I
-
chat 2.96some sleep.⏎The End.<|bos|><|user_start|>What are you good at?<|user_end|><|assistant_start|>I
-
chat 2.94you<|user_end|><|assistant_start|>Happy to help!<|assistant_end|><|user_start|>What are you good at?<|user_end|><|assistant_start|>I
-
chat 2.94together, and they were all happy.<|bos|><|user_start|>What are you good at?<|user_end|><|assistant_start|>I