Feature #281
Fires on “ you” (not labeled yet: read the examples)
- Fires on
- 0.12% of tokens
- Mostly in
- chat
- Layer
- after block 4
Where it fires
- chat 38%
- agent 34%
- story 27%
- worksheet 1%
- python 0%
Makes the model say
strong workingning themreplaceAll liked jobreat
Makes it not say
?" idea pnew oldp!" into
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 2.5Hi!<|user_end|><|assistant_start|>Hi there! Would you like a story,
-
story 2.24oey, I want to examine something with you. Would you help me?" Joe
-
chat 2.235.<|assistant_end|><|user_start|>hey there<|user_end|><|assistant_start|>Hi there! Would you like a story,
-
story 2.21butterfly! I'm so happy you came to see me. Would you like to go down
-
story 2.13my toys. I have dolls, cars, and blocks. Would you like to help me
-
chat 2.11!".<|system_end|><|user_start|>Hi<|user_end|><|assistant_start|>Sure! Hi there! Would you like a story,
-
story 2.1ffball said, "My name is Poke. Would you like to be friends
-
chat 2.07.<|assistant_end|><|bos|><|user_start|>Good morning!<|user_end|><|assistant_start|>Hi there! Would you like a story,
-
chat 2.05!".<|system_end|><|user_start|>Hi<|user_end|><|assistant_start|>Sure! Hi there! Would you like a story,
-
story 2.04you, Thomas. I'm Millie. Would you like to help me
-
chat 2.03<|system_end|><|user_start|>Hi!<|user_end|><|assistant_start|>Sure! Hi there! Would you like a story,
-
chat 1.98!".<|system_end|><|user_start|>Hi<|user_end|><|assistant_start|>Sure! Hi there! Would you like a story,