Feature #3738
Fires on “You” (not labeled yet: read the examples)
- Fires on
- 0.71% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 56%
- chat 33%
- python 7%
- worksheet 3%
- story 1%
Makes the model say
upon listassert prelude=**/<|system_start|> example sure
Makes it not say
was saw6 wanted am went clapped down
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 2.56project.<|assistant_end|><|bos|><|system_start|>tools: calculator<|system_end|><|user_start|>Thank you<|user_end|><|assistant_start|>You're welcome!<|assistant_end|>
-
chat 2.55laptop by mini-lab.<|assistant_end|><|bos|><|user_start|>thanks<|user_end|><|assistant_start|>You're welcome!<|assistant_end|>
-
chat 2.538.<|assistant_end|><|bos|><|system_start|>tools: calculator<|system_end|><|user_start|>Thank you<|user_end|><|assistant_start|>You're welcome!<|assistant_end|>
-
worksheet 2.42trained from scratch on a laptop by mini-lab.<|assistant_end|><|bos|>Could you tell me what
-
chat 2.4to his new friend and went home to get warm.<|bos|><|user_start|>Could you tell me what
-
chat 2.33<|assistant_start|>Happy to help!<|assistant_end|><|bos|><|user_start|>thanks<|user_end|><|assistant_start|>You're welcome!<|assistant_end|>
-
chat 2.29<|system_start|>tools: calculator<|system_end|><|user_start|>How can you help me?<|user_end|><|assistant_start|>I'm prelude, and
-
chat 2.26forest.<|bos|><|system_start|>tools: calculator<|system_end|><|user_start|>thanks<|user_end|><|assistant_start|>You're welcome!<|assistant_end|>
-
chat 2.26⏎```⏎<|bos|><|system_start|>Answer with the number only.<|system_end|><|user_start|>Could you tell me what
-
chat 2.25be happy and fast too.<|bos|><|system_start|>tools: calculator<|system_end|><|user_start|>Could you tell me what
-
chat 2.24Try asking me for a story!<|assistant_end|><|user_start|>thanks<|user_end|><|assistant_start|>You're welcome!<|assistant_end|>
-
chat 2.24<|user_end|><|assistant_start|>Hello! How can I help you today?<|assistant_end|><|user_start|>Could you tell me what