AnonyRepo/dllm-prm-snapshots-gsm8k

该数据集名为dLLM PRM Snapshots on GSM8K (Dream-7B)，是一个基于GSM8K训练问题的中间扩散状态数据集。它使用Dream-7B模型进行策略采样，生成了1,012,992个部分掩码的状态。数据包括4个样本×7,473个问题×约20个快照/轨迹，每个状态带有从最终答案传播的二进制is_correct标签（正率约46%）。采样器采用Dream-7B默认设置，T=128，每6步采样一次。数据集适用于文本分类和其他任务，专注于过程奖励模型和离散扩散研究。

This dataset, named dLLM PRM Snapshots on GSM8K (Dream-7B), consists of on-policy intermediate diffusion states sampled from the Dream-7B model on GSM8K train problems. It includes 1,012,992 partially-masked states, with 4 samples × 7,473 problems × approximately 20 snapshots per trajectory. Each state is labeled with a binary is_correct label propagated from the final answer (positive rate ~46%). The sampler uses Dream-7B defaults, T=128, sampling every 6 steps. It is designed for text-classification and other tasks, focusing on process-reward-model and discrete-diffusion applications.