Feature #2705
Fires on “ you” (not labeled yet: read the examples)
- Fires on
- 0.80% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 94%
- chat 4%
- story 2%
- python 0%
- worksheet 0%
Makes the model say
fromoldString<|assistant_end|> function<|tool_call_start|>That feelingassert
Makes it not say
returnopy best add testatne
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 1.89Hi!<|user_end|><|assistant_start|>Hi there! Would you like a story,
-
story 1.79no! There was no
-
story 1.75. We wanted to explore and look at the pretty flowers. Would you like to join us
-
story 1.74oey, I want to examine something with you. Would you help me?" Joe
-
chat 1.695.<|assistant_end|><|user_start|>hey there<|user_end|><|assistant_start|>Hi there! Would you like a story,
-
agent 1.67pattern=**/test_*.py<|tool_call_end|><|assistant_end|><|tool_start|>test_strings.py<|tool_end|><|assistant_start|><|tool_call_start|>shell
-
chat 1.65add numbers.<|assistant_end|><|user_start|>Hello!<|user_end|><|assistant_start|>Hi there! Would you like a story,
-
story 1.64you, Thomas. I'm Millie. Would you like to help me
-
chat 1.64mini-lab.<|assistant_end|><|user_start|>Hello<|user_end|><|assistant_start|>Hi there! Would you like a story,
-
agent 1.64pattern=**/test_*.py<|tool_call_end|><|assistant_end|><|tool_start|>test_words.py<|tool_end|><|assistant_start|><|tool_call_start|>shell
-
chat 1.64<|assistant_end|><|user_start|>Hi!<|user_end|><|assistant_start|>Hi there! Would you like a story,
-
chat 1.59<|assistant_start|>I'm sorry, I can't help with that. I only know how to tell stories