Feature #2234
Fires on “ sub” (not labeled yet: read the examples)
- Fires on
- 0.44% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 84%
- story 11%
- chat 3%
- python 2%
- worksheet 1%
Makes the model say
lver nows live smallightc
Makes it not say
=assert happeneddefpython<|arg|> upon tell
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
agent 2.95ob, grep, question, read, shell, skill, subagent, we
-
agent 2.93ob, grep, question, read, shell, skill, subagent, we
-
agent 2.93ob, grep, question, read, shell, skill, subagent, we
-
agent 2.93ob, grep, question, read, shell, skill, subagent, we
-
agent 2.92ob, grep, question, read, shell, skill, subagent, we
-
agent 2.92ob, grep, question, read, shell, skill, subagent, we
-
agent 2.91ob, grep, question, read, shell, skill, subagent, we
-
agent 2.91ob, grep, question, read, shell, skill, subagent, we
-
agent 2.91ob, grep, question, read, shell, skill, subagent, we
-
agent 2.91ob, grep, question, read, shell, skill, subagent, we
-
agent 2.91ob, grep, question, read, shell, skill, subagent, we
-
agent 2.9ob, grep, question, read, shell, skill, subagent, we