Feature #480
Fires on “ if” (not labeled yet: read the examples)
- Fires on
- 0.79% of tokens
- Mostly in
- story
- Layer
- after block 4
Where it fires
- story 54%
- agent 30%
- python 10%
- chat 5%
- worksheet 2%
Makes the model say
fl just tinyP e preludebackwards[
Makes it not say
assert threwHey story remember upon "rement
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
story 3.18even more fun.<|bos|>
-
chat 3.16⏎They played and had fun in the sand. Tim learned that even if something looks stupid,
-
story 2.89Tom learned that sometimes, even when things seem bad,
-
story 2.87was overwhelmed with joy. She realised that even if you don't always
-
chat 2.82⏎They played and had fun in the sand. Tim learned that even if something looks stupid,
-
chat 2.8⏎They played and had fun in the sand. Tim learned that even if something looks stupid,
-
story 2.74scared of everything. One
-
story 2.74scared. I'm just
-
story 2.74scared and it started to
-
story 2.68from that day on, Spot and Tim always helped each other when things were hard.<|bos|><|system_start|>
-
story 2.55little cat and the big dog became best friends. They learned that even if someone looks different,
-
chat 2.37avy boat. They sailed the sea and had many fun adventures. Even though Tom could not hear