We use the official KorQuAD 1.0 evaluation script to evaluate our KorQuAD predictions (more accurately than with the HuggingFace SQuAD evaluation script).
Credit: https://korquad.github.io/KorQuad%201.0/
Usage:
python evaluate.py \
--dataset_file </path/to/qa/ko/dev-v1.1.json> \
--prediction_file </path/to/predictions_.json>