Feature #329
Fires on āeā (not labeled yet: read the examples)
- Fires on
- 0.72% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 88%
- story 9%
- python 1%
- chat 1%
- worksheet 0%
Makes the model say
veripke through newcountreas
Makes it not say
:s))ā own was realized real
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
agent 2.93bfetch, websearch, write, execute<|system_end|><|system_start|>
-
agent 2.9bfetch, websearch, write, execute<|system_end|><|system_start|>
-
agent 2.9bfetch, websearch, write, execute<|system_end|><|system_start|>
-
agent 2.9bfetch, websearch, write, execute<|system_end|><|system_start|>
-
agent 2.9bfetch, websearch, write, execute<|system_end|><|system_start|>
-
agent 2.9bfetch, websearch, write, execute<|system_end|><|system_start|>
-
agent 2.86bfetch, websearch, write, execute<|system_end|><|system_start|>
-
agent 2.86bfetch, websearch, write, execute<|system_end|><|system_start|>
-
agent 2.86bfetch, websearch, write, execute<|system_end|><|system_start|>
-
agent 2.86bfetch, websearch, write, execute<|system_end|><|system_start|>
-
agent 2.86bfetch, websearch, write, execute<|system_end|><|system_start|>
-
agent 2.86bfetch, websearch, write, execute<|system_end|><|system_start|>