Feature #552
Fires on “<|user_start|>” (not labeled yet: read the examples)
- Fires on
- 0.76% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 39%
- chat 23%
- story 22%
- python 14%
- worksheet 2%
Makes the model say
s plusakingingsjoinite usant
Makes it not say
Wow lovedread liked's table done be
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
story 2.21would be hard, but they wanted to try.⏎The ants worked as a team to carry
-
story 1.92<|bos|>Once upon a time, there were two friends called Ant and Bee. One day
-
python 1.73lemon")⏎"n"⏎⏎`whisper(s)` returns s in lower case.
-
python 1.71return s + "!"⏎```⏎⏎`excite` returns s with an exclamation mark
-
chat 1.68story about a cat!<|user_end|><|assistant_start|>One day, a big cat and a little dog were playing
-
chat 1.66<|user_start|>What is 6
-
agent 1.66<|user_start|>Create stats.py
-
agent 1.66<|user_start|>Run the tests in
-
agent 1.66<|user_start|>"Why does main
-
agent 1.66<|user_start|>What does hello.
-
chat 1.66<|user_start|>What are you good
-
agent 1.66<|user_start|>Please which functions are