Feature #3255
Fires on âfâ (not labeled yet: read the examples)
- Fires on
- 1.10% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 95%
- python 2%
- chat 2%
- story 1%
- worksheet 0%
Makes the model say
!")ally<|tool_call_end|>`.ap %Run[::
Makes it not say
â s She<|bos|>A watch watched breath
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
agent 2.44, read, shell, skill, subagent, webfetch, we
-
agent 2.43, read, shell, skill, subagent, webfetch, we
-
agent 2.42, read, shell, skill, subagent, webfetch, we
-
agent 2.41, read, shell, skill, subagent, webfetch, we
-
agent 2.4, read, shell, skill, subagent, webfetch, we
-
agent 2.4, read, shell, skill, subagent, webfetch, we
-
agent 2.4, read, shell, skill, subagent, webfetch, we
-
agent 2.39, read, shell, skill, subagent, webfetch, we
-
agent 2.39, read, shell, skill, subagent, webfetch, we
-
agent 2.39, read, shell, skill, subagent, webfetch, we
-
agent 2.39, read, shell, skill, subagent, webfetch, we
-
agent 2.38, read, shell, skill, subagent, webfetch, we