Feature #3182
Fires on โ?โ (not labeled yet: read the examples)
- Fires on
- 0.55% of tokens
- Mostly in
- story
- Layer
- after block 4
Where it fires
- story 42%
- agent 34%
- chat 14%
- python 6%
- worksheet 3%
Makes the model say
isying x`: our<|think_start|>No")`
Makes it not say
friendspython goodbyel strongSoreature
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
agent 3.08?<|user_end|><|assistant_start|><|tool_call_start|>s
-
chat 3.08?<|assistant_end|><|user_start|>Hello!
-
worksheet 3.08?โ171
-
chat 3.08?<|user_end|><|assistant_start|>Once upon
-
chat 3.08?<|user_end|><|assistant_start|>I'm
-
chat 3.08?<|user_end|><|assistant_start|>The answer
-
chat 3.08?<|user_end|><|assistant_start|><|tool_call_start|>calculator
-
chat 3.08?<|user_end|><|assistant_start|>The answer
-
chat 3.08?<|user_end|><|assistant_start|><|think_start|>1
-
chat 3.08?<|user_end|><|assistant_start|><|tool_call_start|>calculator
-
chat 3.08?<|user_end|><|assistant_start|>Once upon
-
story 2.9! The cat was not