CoPE (Context vs. Parametric Evaluation) is a framework for analyzing how large language models use contextual input versus parametric knowledge when answering open-ended questions. It provides tools and evaluation setups for testing knowledge reliance in multilingual settings.
The repository includes the MultiWikiAtomic dataset, which breaks down Wikipedia articles in English, Spanish and Danish into atomic knowledge statements for controlled evaluation.
The CoPE framework and dataset will be available soon.