Feature #90
Fires on “et” (not labeled yet: read the examples)
- Fires on
- 0.78% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 81%
- story 13%
- chat 5%
- worksheet 1%
- python 1%
Makes the model say
s learn makera't helpw,
Makes it not say
main playedStjson = sat playingpped
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
story 2.09Mummy, but he still felt a little frustrated that the lemon hadn't fit in the
-
chat 2.0352.<|assistant_end|><|bos|><|user_start|>Can you book a table for two tonight?<|user_end|><|assistant_start|>
-
agent 1.99fetch, web
-
agent 1.99fetch, web
-
agent 1.99fetch, web
-
agent 1.99fetch, web
-
agent 1.99fetch, web
-
agent 1.99fetch, web
-
story 1.95drink water every day.
-
chat 1.94at the festival.<|bos|><|user_start|>Can you book a table for two tonight?<|user_end|><|assistant_start|>
-
story 1.93She started tracing the letters of the alphabet with the pen.
-
agent 1.91lwoer_case