Feature #698
Fires on โ โ (not labeled yet: read the examples)
- Fires on
- 0.80% of tokens
- Mostly in
- python
- Layer
- after block 4
Where it fires
- python 51%
- chat 14%
- worksheet 13%
- agent 12%
- story 11%
Makes the model say
SueOnceSue Lily add Herup<|assistant_start|>
Makes it not say
forever We away smiledfromWe wemain
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 2.11in the park. He was jumping from
-
chat 1.96was a grumpy boat. The boat was always
-
story 1.96and the cow had a race. They wanted to see who was the fastest.โThe
-
chat 1.94, there was a big dragon. The dragon was very
-
story 1.87Red, "Let's have a race to see who is the fastest!" Red
-
chat 1.83<|system_start|>Do not use the calculator.<|system_end|><|user_start|>What is the sum of 3 and
-
chat 1.81a little girl named Lucy went to the park with her parent. The sun was bright
-
chat 1.78red shirt and blue pants.<|assistant_end|><|user_start|>What is the sum of 4 and
-
story 1.78reddy wanted to become
-
agent 1.78rep<|arg|>pattern=
-
chat 1.76how to tell stories and add numbers.<|assistant_end|><|bos|><|user_start|>What is the sum of 17
-
chat 1.75The answer is 1420.<|assistant_end|><|user_start|>What is the sum of 70