Improved German translations of the ARC-Easy and ARC-Challenge benchmarks (Clark et al., 2018), part of the Boldt German Evaluation Suite.
The validation splits of the datasets were translated from the English originals using Tower+ 72B. Rather than translating question and answer options separately, we translated complete question-answer instances end-to-end to preserve internal consistency and task logic. A small… See the full description on the dataset page:
https://huggingface.co/datasets/Boldt/arc_de.
