Feature #1701
Fires on “<|user_start|>” (not labeled yet: read the examples)
- Fires on
- 2.47% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 49%
- chat 48%
- python 1%
- story 1%
- worksheet 1%
Makes the model say
pl(- counts checks([Added needed
Makes it not say
, that not the in. all sad
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 3.09for a story!<|assistant_end|><|user_start|>What's your name
-
chat 3.01<|user_start|>Bye<|user_end|><|assistant_start|>Goodbye! Have a nice day.<|assistant_end|><|user_start|>what is your name
-
chat 2.980 + 4?<|user_end|><|assistant_start|>The answer is 4.<|assistant_end|><|user_start|>Now add 2
-
chat 2.94. He stumbled on the grass as he walked.<|assistant_end|><|user_start|>Calculate 22
-
chat 2.917 to that?<|user_end|><|assistant_start|>The answer is 41.<|assistant_end|><|user_start|>Now add 3
-
chat 2.91add 2?<|user_end|><|assistant_start|>The answer is 24.<|assistant_end|><|user_start|>Plus 1?
-
chat 2.9night!<|user_end|><|assistant_start|>Bye! Come back for another story soon.<|assistant_end|><|user_start|>write a C
-
chat 2.9He looked at his mom and dad flying high in the sky.<|assistant_end|><|user_start|>Hey, what's
-
chat 2.89<|user_start|>0 + 4<|user_end|><|assistant_start|>The answer is 4.<|assistant_end|><|user_start|>Plus 25
-
chat 2.89night!<|user_end|><|assistant_start|>Goodbye! Have a nice day.<|assistant_end|><|user_start|>Hey, what's
-
chat 2.89night!<|user_end|><|assistant_start|>Bye! Come back for another story soon.<|assistant_end|><|user_start|>Compute 39
-
chat 2.8724.<|user_end|><|assistant_start|>The answer is 25.<|assistant_end|><|user_start|>What if you add