Feature #3491
Fires on “ good” (not labeled yet: read the examples)
- Fires on
- 0.18% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 61%
- story 31%
- chat 6%
- python 1%
- worksheet 0%
Makes the model say
sum heavy sumill lotThisour⏎⏎
Makes it not say
7'll?averageON ever6
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
story 4.05ed in jumping over the wall. They were both very happy and became good friends. And they
-
chat 3.9Tim and the doctor played with toy cars and had fun. They became good friends. Tim was
-
chat 3.87They had lots of fun. The pink bunny and the little girl became good friends. They played
-
story 3.8each had a ball to play with. They both were happy and became good friends.<|bos|>Once
-
story 3.75ran to catch it. They played together all day, and they became best friends. Tim and
-
story 3.71could open his beak again. He thanked Squeaky and they became good friends. From that
-
story 3.68Kitty and Doggy played with the red ball.⏎Kitty and Doggy became good friends. They played
-
story 3.66!" They played baseball together all day. Spot and Tim became best friends. From that
-
story 3.66ran and jumped in the grass.⏎Lily and Ziggy became best friends. Every day
-
chat 3.65angry and bathed in the sun together. They became good friends and shared the
-
story 3.64cat learned to share and play together, and they became good friends.<|bos|><|system_start|>
-
story 3.6Sue was able to drink water and thanked Tim.⏎They both became good friends and played together