Feature #2701
Fires on “c” (not labeled yet: read the examples)
- Fires on
- 1.05% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 59%
- story 19%
- python 14%
- chat 5%
- worksheet 2%
Makes the model say
commandname<|arg|>):alsescript()`.owed
Makes it not say
It( se returns to_She1
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
agent 3.5cute<|system_end|><|system_start|>You
-
agent 3.5cute<|system_end|><|system_start|>You
-
story 3.5cing and rolling
-
story 3.5cense.⏎The
-
agent 3.5cute<|system_end|><|system_start|>You
-
agent 3.5cute<|system_end|><|system_start|>You
-
agent 3.5cute<|system_end|><|system_start|>You
-
agent 3.15s[-1]<|tool_end|>
-
python 3.15s`.⏎⏎```
-
python 3.15s.lower()`.⏎
-
agent 3.15s[::-1]<|tool_end|>
-
agent 3.15s + "!"<|tool_call_end|><|assistant_end|>