Feature #404
Fires on “ar” (not labeled yet: read the examples)
- Fires on
- 0.72% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 75%
- story 18%
- chat 4%
- python 2%
- worksheet 1%
Makes the model say
wasmes's whoen loves am
Makes it not say
upon mud listfrom cleaned promised butupper
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
agent 5.45bsearch, write,
-
agent 5.45bsearch, write,
-
agent 5.45bsearch, write,
-
agent 5.45bsearch, write,
-
agent 4.95arch, write,
-
agent 4.95arch, write,
-
agent 4.95arch, write,
-
agent 4.95arch, write,
-
story 4.46lives. But she
-
agent 4.4search, write,
-
agent 4.4search, write,
-
agent 4.4search, write,