kangaroo_dataset / evaluation_protocol.md
kangaroo-dataset-german's picture
Initial anonymized dataset release
d7d4553 verified

Evaluation Protocol

The benchmark is designed for single-turn exam-mode evaluation.

Each item is presented once with:

  • German problem text,
  • five answer choices A--E,
  • any question image,
  • any image-based answer options.

The model must return either one final answer letter or an explicit abstention token such as Declined.

Scoring follows the original contest rules:

  • correct answer: +points,
  • wrong answer: -points / 4,
  • abstention: 0.

Accuracy treats abstentions as incorrect so that solving performance and abstention policy can be analyzed separately.

Exam-level results should report both raw contest points and normalized percent maximum score:

percent_max = total_points / maximum_possible_points * 100

Evaluation should avoid tool use, retrieval, multi-turn correction, or item revisiting unless explicitly reported as a separate protocol.