Feature #2278
Fires on “ and” (not labeled yet: read the examples)
- Fires on
- 0.26% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 44%
- story 35%
- python 9%
- chat 8%
- worksheet 4%
Makes the model say
see⏎ says ` ? breatht
Makes it not say
short){ > sadisH could
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
story 5.3and hugged Lily. They
-
chat 5.3and add numbers.<|assistant_end|>
-
worksheet 5.3and 933
-
story 5.3and steam filled the
-
chat 5.3and 8.<|user_end|>
-
story 5.3and said, "I
-
story 5.3and walked, showing
-
story 5.3and followed life. They
-
story 5.3and his mom were done
-
story 5.3and brought back to the
-
story 5.3and help mom. And
-
story 5.3and put it back in