Feature #3467
Fires on “ch” (not labeled yet: read the examples)
- Fires on
- 0.33% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 92%
- story 6%
- chat 1%
- worksheet 0%
- python 0%
Makes the model say
MThank f eakl<|assistant_end|> back
Makes it not say
treeOne: onew smelleditem living
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
agent 2.41, shell, skill, subagent, webfetch, webse
-
agent 2.38, shell, skill, subagent, webfetch, webse
-
agent 2.38, shell, skill, subagent, webfetch, webse
-
agent 2.37, shell, skill, subagent, webfetch, webse
-
agent 2.37, shell, skill, subagent, webfetch, webse
-
agent 2.37, shell, skill, subagent, webfetch, webse
-
agent 2.36, shell, skill, subagent, webfetch, webse
-
agent 2.36, shell, skill, subagent, webfetch, webse
-
agent 2.36, shell, skill, subagent, webfetch, webse
-
agent 2.36, shell, skill, subagent, webfetch, webse
-
agent 2.36, shell, skill, subagent, webfetch, webse
-
agent 2.36, shell, skill, subagent, webfetch, webse