Shashwat20/matter-to-mechanism

Matter to Mechanism 数据集包含 2,645 个专家整理的“问题-假设”对，这些数据来自材料科学和电化学领域的同行评议文献。每个条目将一个精确陈述的研究问题与其对应的科学假设联系起来，并包含逐步的推理链、提出的干预措施、机制原理和目标属性。该数据集是 BatteryHypoBench 的基础，这是一个多维基准，用于评估 AI 在科学假设生成方面的能力，不仅测量表面流畅性，还测量机制深度、推理连贯性和物理合理性。数据集包含 20 个结构化字段，如问题陈述、假设、推理步骤等，并提供了详细的统计信息和示例用法。

Matter to Mechanism contains 2,645 expert-curated problem–hypothesis pairs extracted from peer-reviewed materials science and electrochemistry literature. Each entry links a precisely stated research problem to its corresponding scientific hypothesis, complete with a step-by-step reasoning chain, proposed intervention, mechanistic rationale, and target property. The dataset is the foundation of BatteryHypoBench, a multi-dimensional benchmark for evaluating AI co-scientist systems on scientific hypothesis generation — measuring not surface fluency but mechanistic depth, reasoning coherence, and physical plausibility. The dataset includes 20 structured fields such as problem statement, hypothesis, reasoning steps, etc., and provides detailed statistics and example usage.