Feature #3431
Fires on “Can” (not labeled yet: read the examples)
- Fires on
- 0.45% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 68%
- chat 14%
- worksheet 9%
- story 6%
- python 4%
Makes the model say
0 we bottom x<|think_start|> moral result out
Makes it not say
:read. boy Tim with of that
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 2.69together every day.<|bos|><|system_start|>Answer in one short sentence.<|system_end|><|user_start|>Can you tell me a
-
chat 2.61<|bos|><|system_start|>Answer in one short sentence.<|system_end|><|user_start|>Can you tell me a
-
chat 2.58with the toys.<|bos|><|system_start|>Answer in one short sentence.<|system_end|><|user_start|>Can you tell me a
-
chat 2.57laptop by mini-lab.<|assistant_end|><|bos|><|system_start|>tools: calculator<|system_end|><|user_start|>Can you add 9
-
chat 2.55they always helped each other.<|assistant_end|><|bos|><|system_start|>tools: calculator<|system_end|><|user_start|>Can you book a table
-
chat 2.55pass now.<|assistant_end|><|bos|><|system_start|>Answer in one short sentence.<|system_end|><|user_start|>Can you tell me a
-
chat 2.55583<|assistant_end|><|bos|><|system_start|>Answer with the number only.<|system_end|><|user_start|>Can you add 5
-
chat 2.53has 713 stickers.<|bos|><|system_start|>tools: calculator<|system_end|><|user_start|>Can you tell me a
-
chat 2.52299.<|bos|><|system_start|>Answer in one short sentence.<|system_end|><|user_start|>Can you tell me a
-
chat 2.5296 shells.<|bos|><|system_start|>Answer in one short sentence.<|system_end|><|user_start|>Can you tell me a
-
chat 2.49, and they played together happily.<|bos|><|system_start|>tools: calculator<|system_end|><|user_start|>Can you tell me a
-
chat 2.46ever after.<|assistant_end|><|bos|><|system_start|>Answer in one short sentence.<|system_end|><|user_start|>Can you tell me a