Feature #1373
Fires on “ we” (not labeled yet: read the examples)
- Fires on
- 0.41% of tokens
- Mostly in
- story
- Layer
- after block 4
Where it fires
- story 65%
- chat 19%
- agent 9%
- python 4%
- worksheet 3%
Makes the model say
do ex+ even say did doesver
Makes it not say
equalspt turns longerhttpsirements://alse
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
agent 6.96coding agent harness.<|system_end|><|user_start|>Hello!<|user_end|><|assistant_start|>Hi! What should we do in this project
-
story 6.93and didn't know what to do.⏎Do
-
story 6.92sad and didn't know what to do. But then
-
agent 6.91s, n)`.<|assistant_end|><|user_start|>Hey there<|user_end|><|assistant_start|>Hello! What should we do in this project
-
agent 6.77coding agent harness.<|system_end|><|user_start|>Hey there<|user_end|><|assistant_start|>Hello! What should we do in this project
-
agent 6.59agent harness.<|system_end|><|user_start|>good morning<|user_end|><|assistant_start|>Hey! What should we do in this project
-
story 6.58started to sink! She was so scared and didn't know what to do.⏎Suddenly
-
chat 6.54big, smelly waste. They were sad and didn't know what to do.⏎M
-
agent 6.53coding agent harness.<|system_end|><|user_start|>hi<|user_end|><|assistant_start|>Hi! What should we do in this project
-
story 6.51left to eat. Sam was very sad. He didn't know what to do. He looked
-
agent 6.51harness.<|system_end|><|user_start|>"yo"<|user_end|><|assistant_start|>Hey! What should we do in this project
-
story 6.35She looked everywhere for it. Mia was sad and didn't know what to do. She thought