chengshuaizhao/DataAlchemy

DataAlchemy是一个用于可控研究大型语言模型（LLMs）中链式思维（CoT）推理的符号数据集。每条记录将一个提示与逐步推理轨迹和最终答案配对，生成于DataAlchemy框架下。数据集包含多种基础变换组合和不同的推理深度，能够系统性地探究当测试分布偏离训练分布时CoT行为的变化。数据集旨在用于任务泛化和推理深度泛化等研究。

DataAlchemy is a symbolic dataset for the controlled study of Chain-of-Thought (CoT) reasoning in Large Language Models (LLMs). Each record pairs a prompt with a step-by-step reasoning trace and a final answer, produced under the DataAlchemy framework. The files span multiple compositions of base transformations at varying reasoning depths, enabling systematic probes of how CoT behavior shifts as the test distribution drifts from the training distribution. The dataset is intended for tasks like task generalization and reasoning-depth generalization.