Feature #2982
Fires on “ they” (not labeled yet: read the examples)
- Fires on
- 0.69% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 56%
- story 31%
- python 6%
- worksheet 4%
- chat 4%
Makes the model say
_<|think_end|>oldString)) Python %command⏎
Makes it not say
when hitst wereheo is park
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
story 1.57Tim was not selfish anymore. He shared his sugar with his friends and they all had fun together
-
story 1.47dog became best friends. They learned that even if someone looks different, they can still be a
-
story 1.44and proud, and it knew that it could do anything if it tried hard and practiced.<|bos|>
-
story 1.44. So it quit waving in the wind and just hung there.⏎All the
-
story 1.41not find food. The sta
-
story 1.39said, "I like the rain too, but I don't like getting wet."⏎Bob
-
chat 1.38. Tim was proud and knew that he could do anything if he tried hard.<|assistant_end|><|bos|><|system_start|>
-
story 1.36did not like the bug. She told her friends about it, and they all made a plan
-
story 1.35liked to jump and play near the pond. Ducky warned Froggy to be careful near
-
chat 1.35didn't lose his ball. He said thank you to his friends and they all played together with
-
story 1.31. She realised that even if you don't always win the game, you can still get
-
story 1.28Lucy was happy. From that day on, Lucy always made sure to wipe her teddy