Feature #2368
Fires on “ happy” (not labeled yet: read the examples)
- Fires on
- 0.84% of tokens
- Mostly in
- story
- Layer
- after block 4
Where it fires
- story 84%
- chat 10%
- agent 5%
- python 1%
- worksheet 0%
Makes the model say
again happy too and in now at.
Makes it not say
assert trainednumber#}e divisible One
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
story 3.59loved it. Baby brother went back to sleep, and everyone was happy.<|bos|>Once upon
-
story 3.57good time at the party. They ate and laughed. The table felt happy too. It knew
-
story 3.52dog." The food lady laughed and gave Tom his fish. Tom was happy and ate his fish
-
story 3.49friends and they all had fun together. And the magic sugar fairy was happy that Tim had learned
-
story 3.48thanked Tim for his help and gave him a big hug. Tim felt happy and not messy anymore
-
chat 3.47bird was free. The bird was happy. The compassionate man
-
story 3.38said Tim. The bird came and ate some sandwich. The bird was happy.⏎But then
-
story 3.37proud of them. They thanked her and ate their dinner. They were happy and full. They
-
story 3.37ament stayed safe and shiny on the tree. And they all lived happily ever after.<|bos|>
-
chat 3.35put it in the trash. The park was now clean and everyone was happy. Rex and the
-
story 3.32warnings from reliable friends like Mr. Bear. And they lived happily and safely ever
-
story 3.31played and ate together, just like a family. And they all lived happily ever after.<|bos|>