Feature #4093
Fires on “<|user_end|>” (not labeled yet: read the examples)
- Fires on
- 0.85% of tokens
- Mostly in
- chat
- Layer
- after block 4
Where it fires
- chat 53%
- agent 37%
- story 4%
- python 4%
- worksheet 2%
Makes the model say
return{Answer$mdHello Inside item
Makes it not say
thought words Tom Lily thated. decided
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 2.59<|user_end|><|assistant_start|><|think_start|>9+
-
chat 2.59<|user_end|><|assistant_start|>Once upon a
-
agent 2.59<|user_end|><|assistant_start|><|tool_call_start|>gl
-
chat 2.59<|user_end|><|assistant_start|><|think_start|>00
-
agent 2.59<|user_end|><|assistant_start|><|tool_call_start|>gl
-
agent 2.59<|user_end|><|assistant_start|><|tool_call_start|>gre
-
chat 2.59<|user_end|><|assistant_start|><|think_start|>40
-
chat 2.59<|user_end|><|assistant_start|><|think_start|>7+
-
chat 2.59<|user_end|><|assistant_start|>The answer is
-
agent 2.59<|user_end|><|assistant_start|><|tool_call_start|>she
-
agent 2.59<|user_end|><|assistant_start|>Hello! How
-
agent 2.59<|user_end|><|assistant_start|><|tool_call_start|>she