Feature #991
Fires on “e” (not labeled yet: read the examples)
- Fires on
- 0.71% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 89%
- story 6%
- chat 3%
- worksheet 2%
- python 0%
Makes the model say
ora`.>>> exampleillat cut
Makes it not say
= tests day[::- nums way<|assistant_end|>))
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
agent 2.33bfetch, websearch, write, execute<|system_end|><|system_start|>
-
agent 2.29bfetch, websearch, write, execute<|system_end|><|system_start|>
-
agent 2.29bfetch, websearch, write, execute, to
-
agent 2.29bfetch, websearch, write, execute<|system_end|><|system_start|>
-
agent 2.28bfetch, websearch, write, execute<|system_end|><|system_start|>
-
agent 2.28bfetch, websearch, write, execute<|system_end|><|system_start|>
-
agent 2.27bfetch, websearch, write, execute<|system_end|><|system_start|>
-
agent 2.26bfetch, websearch, write, execute<|system_end|><|system_start|>
-
agent 2.26bfetch, websearch, write, execute<|system_end|><|system_start|>
-
agent 2.25bfetch, websearch, write, execute<|system_end|><|system_start|>
-
agent 2.25bfetch, websearch, write, execute<|system_end|><|system_start|>
-
agent 2.25bfetch, websearch, write, execute<|system_end|><|system_start|>