Feature #1034
Fires on โ<|think_start|>โ (not labeled yet: read the examples)
- Fires on
- 0.68% of tokens
- Mostly in
- chat
- Layer
- after block 4
Where it fires
- chat 57%
- worksheet 26%
- agent 12%
- story 4%
- python 0%
Makes the model say
caseice downto5shout way2
Makes it not say
iโ And will also hide followed stories
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 4.15I need to know 19 + 62.<|user_end|><|assistant_start|><|think_start|>19+6
-
chat 4.14much is 994 + 697?<|user_end|><|assistant_start|><|think_start|>994+
-
chat 3.98.<|assistant_end|><|user_start|>Compute 44+918<|user_end|><|assistant_start|><|think_start|>044+9
-
chat 3.93.<|system_end|><|user_start|>Add 34 and 80.<|user_end|><|assistant_start|><|think_start|>34+8
-
chat 3.9What's 55 + 715?<|user_end|><|assistant_start|><|think_start|>055+7
-
chat 3.85<|bos|><|user_start|>What is 44 plus 41?<|user_end|><|assistant_start|><|think_start|>44+4
-
chat 3.84only.<|system_end|><|user_start|>what is 57+21<|user_end|><|assistant_start|><|think_start|>57+2
-
chat 3.82only.<|system_end|><|user_start|>what is 7+51<|user_end|><|assistant_start|><|think_start|>07+51
-
chat 3.77sum of 19 and 801?<|user_end|><|assistant_start|><|think_start|>019+8
-
chat 3.71<|user_start|>Calculate 48 + 629.<|user_end|><|assistant_start|><|think_start|>048+6
-
chat 3.716.<|assistant_end|><|user_start|>what is 4+18<|user_end|><|assistant_start|><|think_start|>04+18
-
chat 3.71221.<|bos|><|user_start|>84+31=<|user_end|><|assistant_start|><|think_start|>84+3