Feature #1016
Fires on “'t” (not labeled yet: read the examples)
- Fires on
- 0.36% of tokens
- Mostly in
- chat
- Layer
- after block 4
Where it fires
- chat 53%
- agent 33%
- story 13%
- python 0%
- worksheet 0%
Makes the model say
was printsone```One everything "!" hole
Makes it not say
returns over the thisprint amazed namedie
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 2.2I don't know about that.
-
story 2.17's the biggest sunflower I've ever seen!
-
chat 2.1crab said, "I'm not a stupid crab. I just don't talk to everyone.
-
chat 1.88iet<|user_end|><|assistant_start|>Sorry, I don't know about that.
-
chat 1.87capital of Brazil?<|user_end|><|assistant_start|>Sorry, I don't know about that.
-
chat 1.85ed the telephone<|user_end|><|assistant_start|>Sorry, I don't know about that.
-
chat 1.81crab said, "I'm not a stupid crab. I just don't talk to everyone.
-
chat 1.8crab said, "I'm not a stupid crab. I just don't talk to everyone.
-
agent 1.79Sorry, I can't do that: I'm a tiny model that only knows small Python project
-
agent 1.75Sorry, I can't do that: I'm a tiny model that only knows small Python project
-
agent 1.74Sorry, I can't do that: I'm a tiny model that only knows small Python project
-
agent 1.72Sorry, I can't do that: I'm a tiny model that only knows small Python project