WorkspaceBench

Open-source evaluation suite that scores how accurately activation-to-text interpretability tools read the intermediate variables a language model holds during a forward pass, including a hallucination-focused eval.