Feature #828
Fires on “<|assistant_start|>” (not labeled yet: read the examples)
- Fires on
- 0.80% of tokens
- Mostly in
- chat
- Layer
- after block 4
Where it fires
- chat 61%
- agent 36%
- worksheet 2%
- story 1%
- python 0%
Makes the model say
<|think_start|> preludeSetpatternoldStringpathFixed<|tool_call_start|>
Makes it not say
b231 bigc64
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 4.43I need to know 92 + 442.<|user_end|><|assistant_start|><|tool_call_start|>calculator<|arg|>expression
-
chat 4.42need to know 615 + 347.<|user_end|><|assistant_start|>The answer is
-
chat 4.29<|user_start|>I need to know 19 + 62.<|user_end|><|assistant_start|><|think_start|>19+
-
chat 4.23I need to know 50 + 454.<|user_end|><|assistant_start|><|think_start|>050
-
chat 4.21<|user_start|>I need to know 91 + 63.<|user_end|><|assistant_start|>The answer is
-
chat 4.2<|user_start|>I need to know 38 + 65.<|user_end|><|assistant_start|>The answer is
-
chat 4.17.<|system_end|><|user_start|>Add 96 and 667.<|user_end|><|assistant_start|><|think_start|>096
-
chat 4.16I need to know 52 + 634.<|user_end|><|assistant_start|>The answer is
-
chat 4.15<|user_start|>I need to know 0 + 508.<|user_end|><|assistant_start|><|think_start|>000
-
chat 4.11<|user_start|>I need to know 75 + 73.<|user_end|><|assistant_start|><|think_start|>75+
-
chat 4.11only.<|system_end|><|user_start|>Calculate 43 + 71.<|user_end|><|assistant_start|><|think_start|>43+
-
chat 4.1<|user_start|>Calculate 7 + 3.<|user_end|><|assistant_start|><|think_start|>7+3