Feature #1970
Fires on “ shell” (not labeled yet: read the examples)
- Fires on
- 0.56% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 86%
- story 9%
- chat 3%
- python 2%
- worksheet 0%
Makes the model say
right swim stuck around all best numbers swam
Makes it not say
They returns One " end == saysAfter
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
agent 2.51tools: edit, glob, grep, question, read, shell, skill,
-
agent 2.5tools: edit, glob, grep, question, read, shell, skill,
-
agent 2.48tools: edit, glob, grep, question, read, shell, skill,
-
agent 2.48tools: edit, glob, grep, question, read, shell, skill,
-
agent 2.47tools: edit, glob, grep, question, read, shell, skill,
-
agent 2.46tools: edit, glob, grep, question, read, shell, skill,
-
agent 2.45tools: edit, glob, grep, question, read, shell, skill,
-
agent 2.44tools: edit, glob, grep, question, read, shell, skill,
-
agent 2.44tools: edit, glob, grep, question, read, shell, skill,
-
agent 2.44tools: edit, glob, grep, question, read, shell, skill,
-
agent 2.44tools: edit, glob, grep, question, read, shell, skill,
-
agent 2.44tools: edit, glob, grep, question, read, shell, skill,