Feature #3009
Fires on “ sub” (not labeled yet: read the examples)
- Fires on
- 0.38% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 85%
- story 9%
- chat 5%
- python 1%
- worksheet 0%
Makes the model say
lyone<|bos|>.river andomeine
Makes it not say
<|arg|> happy was is sad did dark breath
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
agent 2.49ob, grep, question, read, shell, skill, subagent, we
-
agent 2.45ob, grep, question, read, shell, skill, subagent, we
-
agent 2.45ob, grep, question, read, shell, skill, subagent, we
-
agent 2.44ob, grep, question, read, shell, skill, subagent, we
-
agent 2.44ob, grep, question, read, shell, skill, subagent, we
-
agent 2.44ob, grep, question, read, shell, skill, subagent, we
-
agent 2.44ob, grep, question, read, shell, skill, subagent, we
-
agent 2.44ob, grep, question, read, shell, skill, subagent, we
-
agent 2.43ob, grep, question, read, shell, skill, subagent, we
-
agent 2.42ob, grep, question, read, shell, skill, subagent, we
-
agent 2.42ob, grep, question, read, shell, skill, subagent, we
-
agent 2.42ob, grep, question, read, shell, skill, subagent, we