Feature #1915
Fires on “ How” (not labeled yet: read the examples)
- Fires on
- 0.22% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 35%
- chat 31%
- worksheet 14%
- python 12%
- story 9%
Makes the model say
canar today we to!u?"
Makes it not say
utilsecteddeexpressionselves functionsakeopencode
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 3.45<|system_start|>You are a helpful assistant.<|system_end|><|user_start|>Hello<|user_end|><|assistant_start|>Hello! How can I help you
-
chat 3.32tell short stories and add numbers.<|assistant_end|><|user_start|>Hi<|user_end|><|assistant_start|>Hello! How can I help you
-
chat 3.31or with a calculator.<|assistant_end|><|user_start|>Good morning!<|user_end|><|assistant_start|>Hello! How can I help you
-
chat 3.26by step or with a calculator.<|assistant_end|><|user_start|>Hi<|user_end|><|assistant_start|>Hello! How can I help you
-
agent 3.25<|tool_end|><|assistant_start|>All tests pass.<|assistant_end|><|user_start|>Hey there<|user_end|><|assistant_start|>Hi! How can I help with
-
agent 3.25, a coding agent harness.<|system_end|><|user_start|>Hello!<|user_end|><|assistant_start|>Hello! How can I help with
-
chat 3.25Happy to help!<|assistant_end|><|bos|><|user_start|>Hi!<|user_end|><|assistant_start|>Hello! How can I help you
-
chat 3.24returns `"star!"`.⏎<|bos|><|user_start|>Hello!<|user_end|><|assistant_start|>Hello! How can I help you
-
chat 3.231270.<|assistant_end|><|bos|><|user_start|>Hi<|user_end|><|assistant_start|>Hello! How can I help you
-
chat 3.23small and friendly assistant.<|system_end|><|user_start|>hey there<|user_end|><|assistant_start|>Hello! How can I help you
-
agent 3.19, a coding agent harness.<|system_end|><|user_start|>Hey there<|user_end|><|assistant_start|>Hello! How can I help with
-
chat 3.18, or some numbers to add?<|assistant_end|><|user_start|>Hi<|user_end|><|assistant_start|>Hello! How can I help you