Feature #821
Fires on “ each” (not labeled yet: read the examples)
- Fires on
- 0.63% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 53%
- story 29%
- python 14%
- chat 4%
- worksheet 1%
Makes the model say
command smiled Python other smileed<|user_start|> replied
Makes it not say
'?"⏎⏎.ter wallles])
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
story 3.67big tree, where they hid their toys and snacks. They trusted each other and shared everything
-
chat 3.67and Riri were the best of friends, and they always helped each other.<|assistant_end|><|bos|>
-
story 3.63the nuts they found. They were happy to be friends and to help each other.<|bos|>Once
-
story 3.58problems. And from that day on, Spot and Tim always helped each other when things were
-
story 3.57line at the same time. They both looked at each other in amazem
-
story 3.51Chirpy. From that day on, they became best friends and always helped each other.<|bos|><|system_start|>
-
story 3.48letely surrounded, and John and Jane hugged each other. John whis
-
story 3.39on, Sam and Birdy became best friends, and they always helped each other find acorns
-
story 3.38's better to try and do the right thing."⏎They both hugged each other. Then St
-
story 3.35en to learn from the old man, Pete and Tom hugged each other and said sorry
-
story 3.32safely. They promised to always play the right way and to help each other remember to be
-
story 3.32. He wanted to go back to the big space.⏎They saw each other and called out