Feature #975
Fires on “<|think_start|>” (not labeled yet: read the examples)
- Fires on
- 0.45% of tokens
- Mostly in
- chat
- Layer
- after block 4
Where it fires
- chat 62%
- worksheet 26%
- agent 10%
- story 2%
- python 0%
Makes the model say
typo<|think_start|> inside<|tool_call_start|> around laptop firstquiet
Makes it not say
Thank yourdefi They wavePri
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 5.48not use the calculator.<|system_end|><|user_start|>Compute 7+2<|user_end|><|assistant_start|><|think_start|>7+2:
-
chat 5.41.<|system_end|><|user_start|>Add 34 and 80.<|user_end|><|assistant_start|><|think_start|>34+8
-
chat 5.39now.<|assistant_end|><|bos|><|user_start|>Add 9 and 9.<|user_end|><|assistant_start|><|think_start|>9+9:
-
chat 5.37<|user_start|>Calculate 7 + 3.<|user_end|><|assistant_start|><|think_start|>7+3:
-
chat 5.35number only.<|system_end|><|user_start|>Calculate 3 + 6.<|user_end|><|assistant_start|><|think_start|>3+6:
-
chat 5.31.<|system_end|><|user_start|>Add 38 and 62.<|user_end|><|assistant_start|><|think_start|>38+6
-
chat 5.28<|assistant_end|><|bos|><|user_start|>Add 46 and 13.<|user_end|><|assistant_start|><|think_start|>46+1
-
chat 5.28the calculator.<|system_end|><|user_start|>Compute 54+34<|user_end|><|assistant_start|><|think_start|>54+3
-
chat 5.27number only.<|system_end|><|user_start|>Compute 96+74<|user_end|><|assistant_start|><|think_start|>96+7
-
chat 5.27only.<|system_end|><|user_start|>what is 57+21<|user_end|><|assistant_start|><|think_start|>57+2
-
chat 5.25the calculator.<|system_end|><|user_start|>Calculate 7 + 0.<|user_end|><|assistant_start|><|think_start|>7+0:
-
chat 5.25<|user_start|>Can you add 1 and 5 for me?<|user_end|><|assistant_start|><|think_start|>1+5: