ChemBench

NAME = ChemBench
DATASET_PATH = jablonkagroup/ChemBench
SAMPLE_SPLIT = train
FEWSHOT_SPLIT = train
RESPONSE_TYPE = LOGLIKELIHOODS
METRICS = [AccuracyLoglikelihood, AccuracyNormLoglikelihood]
SUBJECTS = ['analytical_chemistry', 'chemical_preference', 'general_chemistry', 'inorganic_chemistry', 'materials_science', 'organic_chemistry', 'physical_chemistry', 'technical_chemistry', 'toxicity_and_safety']
LANGUAGE = <Language.ENG: 'English'>

• Module: eval_framework.tasks.benchmarks.chembench

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

• Link to dataset: https://huggingface.co/datasets/jablonkagroup/ChemBench

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 "ChemBench".