Feature #464
Fires on “'re” (not labeled yet: read the examples)
- Fires on
- 1.48% of tokens
- Mostly in
- chat
- Layer
- after block 4
Where it fires
- chat 37%
- agent 36%
- python 20%
- story 3%
- worksheet 3%
Makes the model say
function<|assistant_start|> typo<|system_end|><|arg|>ually([ calculator
Makes it not say
7 ask on,8 all54
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 4.06")` returns `5`.⏎<|bos|><|user_start|>thanks<|user_end|><|assistant_start|>You're welcome!<|assistant_end|><|bos|>
-
chat 4.02by mini-lab.<|assistant_end|><|bos|><|user_start|>thanks<|user_end|><|assistant_start|>You're welcome!<|assistant_end|><|user_start|>
-
chat 3.98.py⏎all tests passed⏎<|bos|><|user_start|>Thank you<|user_end|><|assistant_start|>You're welcome!<|assistant_end|><|user_start|>
-
chat 3.9.<|assistant_end|><|bos|><|system_start|>tools: calculator<|system_end|><|user_start|>Thank you<|user_end|><|assistant_start|>You're welcome!<|assistant_end|><|user_start|>
-
chat 3.74.<|bos|><|system_start|>tools: calculator<|system_end|><|user_start|>thanks<|user_end|><|assistant_start|>You're welcome!<|assistant_end|><|bos|>
-
chat 3.73is 133.<|assistant_end|><|user_start|>thanks<|user_end|><|assistant_start|>You're welcome!<|assistant_end|><|bos|>
-
chat 3.72.<|assistant_end|><|bos|><|system_start|>tools: calculator<|system_end|><|user_start|>Thank you<|user_end|><|assistant_start|>You're welcome!<|assistant_end|><|user_start|>
-
chat 3.66102.<|assistant_end|><|bos|><|user_start|>thanks<|user_end|><|assistant_start|>You're welcome!<|assistant_end|><|user_start|>
-
chat 3.65Happy to help!<|assistant_end|><|bos|><|user_start|>thanks<|user_end|><|assistant_start|>You're welcome!<|assistant_end|><|user_start|>
-
chat 3.61.<|bos|><|system_start|>tools: calculator<|system_end|><|user_start|>thanks<|user_end|><|assistant_start|>You're welcome!<|assistant_end|><|bos|>
-
chat 3.57py⏎all tests passed⏎<|bos|><|user_start|>Thanks!<|user_end|><|assistant_start|>You're welcome!<|assistant_end|><|user_start|>
-
agent 3.52mean(a, b)`.<|assistant_end|><|user_start|>Thank you!<|user_end|><|assistant_start|>You're welcome!<|assistant_end|><|bos|>