Feature #237
Fires on “ you” (not labeled yet: read the examples)
- Fires on
- 0.23% of tokens
- Mostly in
- worksheet
- Layer
- after block 4
Where it fires
- worksheet 26%
- chat 25%
- python 19%
- agent 16%
- story 14%
Makes the model say
ai lastlast happilyOkayily backides
Makes it not say
th did But about isvedwherese
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 3.5mini-lab.<|assistant_end|><|user_start|>Hello<|user_end|><|assistant_start|>Hi there! Would you like a story,
-
chat 3.45with a calculator.<|assistant_end|><|user_start|>Hi<|user_end|><|assistant_start|>Hi there! Would you like a story,
-
chat 3.39add numbers.<|assistant_end|><|user_start|>Hello!<|user_end|><|assistant_start|>Hi there! Would you like a story,
-
chat 3.37Hi!<|user_end|><|assistant_start|>Hi there! Would you like a story,
-
chat 3.375.<|assistant_end|><|user_start|>hey there<|user_end|><|assistant_start|>Hi there! Would you like a story,
-
chat 3.34!".<|system_end|><|user_start|>Hi<|user_end|><|assistant_start|>Sure! Hi there! Would you like a story,
-
chat 3.34!".<|system_end|><|user_start|>Hello<|user_end|><|assistant_start|>Sure! Hi there! Would you like a story,
-
chat 3.3assistant.<|system_end|><|user_start|>hey there<|user_end|><|assistant_start|>Hi there! Would you like a story,
-
chat 3.29!".<|system_end|><|user_start|>Hi<|user_end|><|assistant_start|>Sure! Hi there! Would you like a story,
-
chat 3.28<|system_end|><|user_start|>Hi!<|user_end|><|assistant_start|>Sure! Hi there! Would you like a story,
-
story 3.26smiled and said, "Thank you for saying hello. Would you like to share my
-
chat 3.25<|system_end|><|user_start|>Hi!<|user_end|><|assistant_start|>Sure! Hi there! Would you like a story,