Feature #562
Fires on โ<|assistant_end|>โ (not labeled yet: read the examples)
- Fires on
- 0.99% of tokens
- Mostly in
- chat
- Layer
- after block 4
Where it fires
- chat 68%
- agent 29%
- story 2%
- python 0%
- worksheet 0%
Makes the model say
<|bos|>โlyred themreplaceAllOnceies
Makes it not say
impress and pretty % dance.def take
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 4.2know about that. I can only tell short stories and add numbers.<|assistant_end|><|bos|>Tom is a
-
chat 4.19know about that. I can only tell short stories and add numbers.<|assistant_end|><|bos|><|system_start|>tools:
-
chat 4.18know about that. I can only tell short stories and add numbers.<|assistant_end|><|bos|>Lila and Sam
-
chat 4.1know about that. I can only tell short stories and add numbers.<|assistant_end|><|bos|>Once upon a
-
chat 4.1know about that. I can only tell short stories and add numbers.<|assistant_end|><|bos|>Bug: `
-
chat 4.06know about that. I can only tell short stories and add numbers.<|assistant_end|><|bos|><|user_start|>Can you
-
chat 4.06know about that. I can only tell short stories and add numbers.<|assistant_end|><|bos|><|system_start|>tools:
-
chat 4.0know about that. I can only tell short stories and add numbers.<|assistant_end|><|user_start|>Can you tell
-
chat 3.98know about that. I can only tell short stories and add numbers.<|assistant_end|><|bos|>Once upon a
-
chat 3.98know about that. I can only tell short stories and add numbers.<|assistant_end|><|user_start|>tell me
-
chat 3.94know about that. I can only tell short stories and add numbers.<|assistant_end|><|user_start|>what is the
-
chat 3.93know about that. I can only tell short stories and add numbers.<|assistant_end|><|bos|>Ben and Lily