Datasets:
Three likely incorrect answer keys in the validation split
While evaluating a model on the validation split I found three items where the labelled answer looks wrong, and where the model's disagreement was confident and, I think, correct. Flagging them in case they're useful — I'm not proposing a change to the labels themselves, since published scores are compared against them.
| id | question | labelled | arguably correct |
|---|---|---|---|
159d50e325b59c6d29ec371500e173b4 |
What is a form of anaerobic exercising? | run | lift weights |
7f25dbab26165b3c8800c2733ca759d6 |
John was an aristocratic fox hunter. Where might he live? | new hampshire | england |
ede4d302fc2ffe07703158f83c1493f2 |
If there is a place that is hot and arid, what could it be? | lifeless | sandy |
Weightlifting is the textbook anaerobic exercise while running is primarily aerobic; aristocratic fox hunting is characteristically English; and "hot and arid" describes a desert, which is sandy rather than lifeless.
Two further items are arguable and I'm not flagging them as errors: "Where would you put a container can after you buy it?" (labelled garage) and "What will happen to animals after eating food?" (labelled feel pleasure).