YefanZhou98/DLM-Decoding-Analysis

DLM-Decoding-Analysis数据集包含了LLaDA-8B-Instruct扩散语言模型在GSM8K和MMLU-STEM基准测试上的逐步解码轨迹。数据集按不同的重新掩码策略和约束条件分为8个文件夹，每个文件夹包含每个问题的PyTorch检查点，存储逐步解码信息。该数据集旨在研究答案出现行为并设计早期退出解码策略。数据集还包括解码配置、文件格式、快速入门指南、统计信息、许可证和引用信息。

The DLM-Decoding-Analysis dataset contains step-by-step decoding trajectories of the LLaDA-8B-Instruct diffusion language model on the GSM8K and MMLU-STEM benchmarks. The dataset is organized into 8 folders based on different remasking policies and constraints, with each folder containing PyTorch checkpoints per question that store step-by-step decoding information. The dataset is designed to study answer-emergence behavior and to design early-exit decoding strategies. It also includes information on decoding configuration, file format, quick start guide, statistics, license, and citation.