Feature #1647
Fires on “ others” (not labeled yet: read the examples)
- Fires on
- 0.13% of tokens
- Mostly in
- story
- Layer
- after block 4
Where it fires
- story 48%
- agent 29%
- python 14%
- chat 8%
- worksheet 2%
Makes the model say
arER:`.ite can firstaringAnswer
Makes it not say
Sorry1 = many anything toyi tiny
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
story 1.84toy stopped growing, and Sue and Tim learned that sharing with others can make everyone happy
-
story 1.84. She learned to be patient and kind and to always share love with others.<|bos|><|system_start|>tools
-
story 1.8share with one another. The moral of the story is that sharing with others brings happiness and
-
story 1.8was and wanted to be like her. They learned that if you help others, you will feel
-
story 1.79the story is that being creative can help you solve problems and help others.<|bos|><|system_start|>Answer
-
story 1.72truly amazed her. There was a beautiful garden, full of inviting flowers.⏎
-
chat 1.7It flew away and soon stopped above a lake. The water looked inviting, so the
-
chat 1.68. The princess was curious, so she ran up the stairs to investigate.
-
story 1.66big red mailbox up ahead. He was curious and walked closer to investigate.
-
story 1.64. The small leaf learned that sometimes you need to let go to help others.<|bos|><|user_start|>Sam
-
agent 1.64defines upper_case?<|user_end|><|assistant_start|>Locate upper_case definition<|assistant_end|><|bos|>
-
story 1.63ceived an invitation in the mail - a very special invitation. It