LabBenchMC_OLMES

NAME = LabBenchMC_OLMES
DATASET_PATH = futurehouse/lab-bench
SAMPLE_SPLIT = train
FEWSHOT_SPLIT = train
RESPONSE_TYPE = LOGLIKELIHOODS
METRICS = [AccuracyLoglikelihood, AccuracyNormLoglikelihood, AccuracyCompletion, BitsPerByteLoglikelihood]
SUBJECTS = ['CloningScenarios', 'DbQA', 'FigQA', 'LitQA2', 'ProtocolQA', 'SeqQA', 'SuppQA', 'TableQA']
LANGUAGE = <Language.ENG: 'English'>

• Module: eval_framework.tasks.benchmarks.lab_bench

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

• Link to dataset: https://huggingface.co/datasets/futurehouse/lab-bench

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 "LabBenchMC_OLMES".