Feature #1902
Fires on “ answer” (not labeled yet: read the examples)
- Fires on
- 0.48% of tokens
- Mostly in
- chat
- Layer
- after block 4
Where it fires
- chat 66%
- agent 16%
- story 9%
- worksheet 8%
- python 1%
Makes the model say
meet latep eachypythona
Makes it not say
nums named met full adds plus * ==
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 3.4<|tool_call_end|><|assistant_end|><|tool_start|>50<|tool_end|><|assistant_start|>The answer is 50
-
chat 3.366<|think_end|>The answer is 11
-
chat 3.36for me?<|user_end|><|assistant_start|>The answer is 11
-
chat 3.34<|tool_call_end|><|assistant_end|><|tool_start|>7<|tool_end|><|assistant_start|>The answer is 7.
-
chat 3.31<|think_end|>The answer is 13
-
chat 3.31<|think_end|>The answer is 32
-
chat 3.31<|tool_call_end|><|assistant_end|><|tool_start|>320<|tool_end|><|assistant_start|>The answer is 32
-
chat 3.28have now?<|user_end|><|assistant_start|>The answer is 12
-
chat 3.27<|user_end|><|assistant_start|>The answer is 9.
-
chat 3.27<|user_end|><|assistant_start|>The answer is 14
-
chat 3.25The answer is 85
-
chat 3.25The answer is 10