Feature #1682
Fires on “ made” (not labeled yet: read the examples)
- Fires on
- 0.44% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 55%
- chat 22%
- story 16%
- python 6%
- worksheet 0%
Makes the model say
There today toys3 left next b made
Makes it not say
]`⏎`def=**/*. Pythonhinglistlib
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 2.04<|user_start|>thanks<|user_end|><|assistant_start|>Happy to help!<|assistant_end|><|user_start|>Who made you?<|user_end|><|assistant_start|>
-
chat 2.0<|user_start|>thanks<|user_end|><|assistant_start|>Happy to help!<|assistant_end|><|user_start|>Who made you?<|user_end|><|assistant_start|>
-
story 1.99cake from the back of the
-
chat 1.98⏎<|bos|><|user_start|>Thank you<|user_end|><|assistant_start|>You're welcome!<|assistant_end|><|user_start|>Who made you?<|user_end|><|assistant_start|>
-
chat 1.89Bye<|user_end|><|assistant_start|>Goodbye! Have a nice day.<|assistant_end|><|user_start|>Who are you?<|user_end|>
-
agent 1.86is not his!<|bos|><|system_start|>You are a title generator.<|system_end|><|user_start|>Run main.py.
-
chat 1.85hello.py?"<|user_end|><|assistant_start|>Run hello.py<|assistant_end|><|bos|><|user_start|>Who made you?<|user_end|><|assistant_start|>
-
chat 1.79<|assistant_start|>Bye! Come back for another story soon.<|assistant_end|><|bos|><|user_start|>Who made you?<|user_end|><|assistant_start|>
-
agent 1.78(items)`.<|assistant_end|><|bos|><|system_start|>You are a title generator.<|system_end|><|user_start|>Run the test suite
-
chat 1.77<|tool_end|><|assistant_start|>The answer is 110.<|assistant_end|><|bos|><|user_start|>Who made you?<|user_end|><|assistant_start|>
-
chat 1.76good friends. They played with the triangle every day.<|assistant_end|><|bos|><|user_start|>Who won the last W
-
chat 1.7378.<|assistant_end|><|user_start|>Who are you?<|user_end|>