Feature #761
Fires on “ that” (not labeled yet: read the examples)
- Fires on
- 0.24% of tokens
- Mostly in
- story
- Layer
- after block 4
Where it fires
- story 72%
- chat 22%
- agent 5%
- worksheet 0%
- python 0%
Makes the model say
.assert every the all ear: listen
Makes it not say
config editest]`notwhereAddednew
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 4.37erlin today?<|user_end|><|assistant_start|>Sorry, I don't know about that. I can only
-
chat 4.22and Juliet?<|user_end|><|assistant_start|>Sorry, I don't know about that. I can only
-
chat 3.942 times 8?<|user_end|><|assistant_start|>Sorry, I don't know about that. I can only
-
chat 3.92Brazil?<|user_end|><|assistant_start|>Sorry, I don't know about that. I can only
-
chat 3.91Brazil.<|user_end|><|assistant_start|>Sorry, I don't know about that. I can only
-
chat 3.89C++ code?<|user_end|><|assistant_start|>Sorry, I don't know about that. I can only
-
chat 3.85French fast?<|user_end|><|assistant_start|>Sorry, I don't know about that. I can only
-
chat 3.83s scores.<|user_end|><|assistant_start|>Sorry, I don't know about that. I can only
-
chat 3.82away is the moon?<|user_end|><|assistant_start|>Sorry, I don't know about that. I can only
-
chat 3.81away is the moon?<|user_end|><|assistant_start|>Sorry, I don't know about that. I can only
-
chat 3.77s scores.<|user_end|><|assistant_start|>Sorry, I don't know about that. I can only
-
chat 3.74is London?<|user_end|><|assistant_start|>Sorry, I don't know about that. I can only