Feature #2781
Fires on ācā (not labeled yet: read the examples)
- Fires on
- 0.49% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 79%
- story 13%
- python 3%
- chat 2%
- worksheet 2%
Makes the model say
hriteiceoldStringpath rever prelude typo
Makes it not say
ā his he " over of in
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
story 4.84went to the park to do some exercise. He loved
-
agent 4.67fetch, websearch, write, execute, conte
-
agent 4.65, write, execute<|system_end|><|system_start|>You
-
agent 4.65, write, execute<|system_end|><|system_start|>You
-
agent 4.65, write, execute<|system_end|><|system_start|>You
-
agent 4.65, write, execute, conte
-
agent 4.65, write, execute<|system_end|><|system_start|>You
-
agent 4.65, write, execute<|system_end|><|system_start|>You
-
agent 4.63bsearch, write, execute<|system_end|><|system_start|>You
-
agent 4.63bsearch, write, execute<|system_end|><|system_start|>You
-
agent 4.63bsearch, write, execute<|system_end|><|system_start|>You
-
agent 4.63bsearch, write, execute<|system_end|><|system_start|>You