Feature #2301
Fires on “ you” (not labeled yet: read the examples)
- Fires on
- 0.38% of tokens
- Mostly in
- chat
- Layer
- after block 4
Where it fires
- chat 62%
- agent 16%
- story 15%
- worksheet 8%
- python 1%
Makes the model say
doing tell mean two need window, do
Makes it not say
was fell The]doubledmin wasn `
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 5.66<|user_start|>What are you good at?<|user_end|>
-
chat 5.63<|bos|><|user_start|>What are you good at?<|user_end|>
-
chat 5.63<|bos|><|user_start|>What are you good at?<|user_end|>
-
chat 5.387<|think_end|>The answer is 17.<|assistant_end|><|bos|><|user_start|>What are you good at?<|user_end|>
-
agent 5.24running in OpenCode, a coding agent harness.<|system_end|><|user_start|>"What are you able to do?"
-
chat 5.1947 + 25 = 72<|bos|><|user_start|>What are you good at?<|user_end|>
-
chat 5.08answer is 15.<|assistant_end|><|user_start|>what are you good at<|user_end|><|assistant_start|>
-
story 5.06walked up to him.⏎"Hey there son. What are you doing?"⏎B
-
chat 5.04story soon.<|assistant_end|><|user_start|>What are you good at?<|user_end|>
-
chat 5.03<|assistant_start|>The answer is 630.<|assistant_end|><|bos|><|user_start|>What are you good at?<|user_end|>
-
agent 5.03return -x`. The tests pass now.<|assistant_end|><|user_start|>"What are you?"<|user_end|><|assistant_start|>I
-
chat 5.02<|think_end|>The answer is 816.<|assistant_end|><|bos|><|user_start|>What are you good at?<|user_end|>