Feature #422
Fires on “ and” (not labeled yet: read the examples)
- Fires on
- 0.13% of tokens
- Mostly in
- python
- Layer
- after block 4
Where it fires
- python 55%
- agent 25%
- story 13%
- chat 4%
- worksheet 2%
Makes the model say
import their=of is lesson a<|bos|>
Makes it not say
trainedles push sorry grass Sorry broke
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
story 2.9and hugged Lily. They
-
chat 2.9and add numbers.<|assistant_end|>
-
worksheet 2.9and 933
-
story 2.9and steam filled the
-
chat 2.9and 8.<|user_end|>
-
story 2.9and said, "I
-
story 2.9and walked, showing
-
story 2.9and followed life. They
-
story 2.9and his mom were done
-
story 2.9and brought back to the
-
story 2.9and help mom. And
-
story 2.9and put it back in