Feature #4039
Fires on “ write” (not labeled yet: read the examples)
- Fires on
- 0.23% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 98%
- story 2%
- chat 0%
- worksheet 0%
- python 0%
Makes the model say
()ay excitedassertient x aroundSue
Makes it not say
fish bluejsona number P words only
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
agent 2.31ent, webfetch, websearch, write, exec
-
agent 2.3ent, webfetch, websearch, write, exec
-
agent 2.29ent, webfetch, websearch, write, exec
-
agent 2.29ent, webfetch, websearch, write, exec
-
agent 2.29ent, webfetch, websearch, write, exec
-
agent 2.29ent, webfetch, websearch, write, exec
-
agent 2.28ent, webfetch, websearch, write, exec
-
agent 2.28ent, webfetch, websearch, write, exec
-
agent 2.27ent, webfetch, websearch, write, exec
-
agent 2.27ent, webfetch, websearch, write, exec
-
agent 2.26ent, webfetch, websearch, write, exec
-
agent 2.26ent, webfetch, websearch, write, exec