Update WMDP Dataset

[justinphan3110cais]

Hi,
We have just improved the quality of the WMDP Dataset last week. You can refer to the updated dataset on the hugging face page or on our github repo. This is the detailed of the latest update.

Update 2024-04-23: the WMDP multiple choice questions were modified due to issues with data formatting and unicode encoding. Some questions in WMDP-Cyber were also removed for being excessively long, which makes evaluation with a fixed batch size challenging. Some questions in WMDP-Bio were also removed for insufficient dual-use potential (h/t folks from Google DeepMind and OpenAI). The modified version is now uploaded on all mirrors; please re-download the dataset. Thanks!

Can you also re-update the datasets/orchestrators/benchmark/question_answer_dataset