Feature #1944
Fires on “ new” (not labeled yet: read the examples)
- Fires on
- 0.26% of tokens
- Mostly in
- story
- Layer
- after block 4
Where it fires
- story 73%
- chat 13%
- agent 9%
- worksheet 3%
- python 2%
Makes the model say
s friend toy friends st onef house
Makes it not say
)` wagged but both>>>Spot]`us
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
agent 4.04new file tools.py
-
story 4.04new to the factory
-
story 4.04new game in the store
-
story 4.04new things together. As
-
chat 3.94they clapped their hands and laughed. They were so happy they made a new friend and trad
-
story 3.88and made a new friend.<|bos|>One
-
story 3.74and play all day long.⏎One day, Ducky met a new friend named Froggy
-
story 3.66said, "Max, let's go to the store and buy a new red ball!" Max
-
story 3.61bathe in the sun every day. One day, Tim met a new friend, a little
-
story 3.61. They went back to their march, happy to have met a new friend.<|bos|><|user_start|>
-
story 3.6me free!" Fred was surprised, but he smiled. He made a new friend, and they
-
story 3.59trains along the tracks.⏎One day, they found a new train in their toy