Feature #2270
Fires on “o” (not labeled yet: read the examples)
- Fires on
- 0.55% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 48%
- story 34%
- worksheet 11%
- chat 6%
- python 1%
Makes the model say
txt sad$ crycommand passes hit I
Makes it not say
toy idea way.ffut next(
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
story 1.77scared and it started to fall
-
agent 1.72⏎ print(vowls("lemon"))⏎ ^^^^
-
story 1.65She sees Tom and the ink. She is not happy. She says, "Tom, what
-
story 1.65didn't want her daddy to be bored. She said, "Let's go
-
agent 1.63<|tool_end|><|assistant_start|>This project has 4 files:
-
story 1.61a pet rabbit, Cotton, and they decided to go on an adventure and explore the
-
chat 1.584+61: 4+1=5, 5
-
python 1.56o":`.⏎⏎
-
story 1.56o turned to Jane and
-
story 1.56oes and saving
-
agent 1.55mean to helpers.py
-
chat 1.54Sam was a little boy.<|assistant_end|><|bos|>Once