Feature #2023
Fires on “ they” (not labeled yet: read the examples)
- Fires on
- 0.11% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 49%
- story 32%
- python 13%
- chat 3%
- worksheet 3%
Makes the model say
<|assistant_end|><|arg|>ate were1 calculator⏎ic
Makes it not say
where sees old small⏎ rub becausemin
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
story 1.62soft and warm and smelled great. Daddy and Mommy shared them and said they were the best eggs
-
story 1.6do a quiz to learn about the things they collected. She said they could not go to
-
story 1.45again. She said guns could hurt them and others. She said they were lucky nothing bad
-
story 1.35soap. She took them inside and gave them a snack. She said they could play in the
-
agent 1.31ery, memory<|system_end|><|system_start|>You are an AI agent
-
story 1.3Stop fighting! You need to learn to be friends. If you work together, you can do
-
story 1.28He said he needed all the blocks
-
story 1.28He said they could pray for themselves and for others. He said they could
-
story 1.25to come down. He said they were naughty and dangerous. He said they could fall and get
-
story 1.25urkey and mashed potatoes. She kissed them and said she was proud of them
-
agent 1.23= 1⏎ for i in range(1, n + 1):⏎ result =
-
agent 1.22<|system_start|>tools: edit, gl