Feature #983
Fires on “ent” (not labeled yet: read the examples)
- Fires on
- 0.95% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 87%
- worksheet 5%
- story 4%
- chat 3%
- python 1%
Makes the model say
<|assistant_end|><|tool_call_start|>scommand wokeed our //
Makes it not say
very so but much2⏎ excited But
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
agent 2.41grep, question, read, shell, skill, subagent, webf
-
agent 2.41grep, question, read, shell, skill, subagent, webf
-
agent 2.4grep, question, read, shell, skill, subagent, webf
-
agent 2.39grep, question, read, shell, skill, subagent, webf
-
agent 2.39grep, question, read, shell, skill, subagent, webf
-
agent 2.38grep, question, read, shell, skill, subagent, webf
-
agent 2.38grep, question, read, shell, skill, subagent, webf
-
agent 2.38grep, question, read, shell, skill, subagent, webf
-
agent 2.37grep, question, read, shell, skill, subagent, webf
-
agent 2.37grep, question, read, shell, skill, subagent, webf
-
agent 2.37grep, question, read, shell, skill, subagent, webf
-
agent 2.37grep, question, read, shell, skill, subagent, webf