ARC

NAME = ARC
DATASET_PATH = ai2_arc
SAMPLE_SPLIT = test
FEWSHOT_SPLIT = train
RESPONSE_TYPE = LOGLIKELIHOODS
METRICS = [AccuracyLoglikelihood, AccuracyNormLoglikelihood]
SUBJECTS = ['ARC-Easy', 'ARC-Challenge']
LANGUAGE = <Language.ENG: 'English'>

• Module: eval_framework.tasks.benchmarks.arc

• File: src/eval_framework/tasks/benchmarks/arc.py | View on GitHub

• Link to dataset: https://huggingface.co/datasets/ai2_arc

More detailed documentation, with prompt examples and ground truth completions, can be generated with uv run -m eval_framework.utils.generate_task_docs --add-prompt-examples --only-tasks "ARC".