SMEPO
收藏资源简介:
该数据集为论文《Hide to Guide: Learning via Semantic Masking》及SMEPO(Semantic Masked Expert Policy Optimization)方法提供支持。SMEPO是一种基于专家指导的强化学习方法,通过掩码专家轨迹中与奖励相关的语义片段,同时保留其过程结构,以改进语言模型。数据集包含数学、代码和智能体搜索等任务的原始专家轨迹,用于指导模型训练。所有数据采用统一的结构模式,包含三个核心字段:question(输入问题或任务描述)、reward_model(验证器或奖励模型信号信息)和teacher_ds(任务特定的专家轨迹)。数据集主要用于文本生成任务,用户可通过提供的脚本从原始专家轨迹构建掩码版本的数据集,以支持SMEPO方法的实验与应用。
This dataset supports the paper Hide to Guide: Learning via Semantic Masking and the SMEPO (Semantic Masked Expert Policy Optimization) method. SMEPO is a reinforcement learning method based on expert guidance, which improves language models by masking reward-related semantic segments in expert trajectories while preserving their process structure. The dataset contains raw expert trajectories for tasks such as mathematics, coding, and agent search, used to guide model training. All data follows a unified structural pattern, including three core fields: question (input problem or task description), reward_model (validator or reward model signal information), and teacher_ds (task-specific expert trajectories). The dataset is primarily used for text generation tasks, and users can construct masked versions from the raw expert trajectories using provided scripts to support experiments and applications of the SMEPO method.
数据集名称
SMEPO (Semantic Masked Expert Policy Optimization)
任务类别
- 文本生成 (text-generation)
数据集概述
该数据集用于论文《Hide to Guide: Learning via Semantic Masking》,属于专家引导的可验证奖励强化学习(RLVR)方法。SMEPO 通过掩码专家轨迹中的奖励相关语义跨度,同时保留其过程结构,从而改进语言模型。数据集包含数学、代码和智能体搜索任务的原始专家轨迹。
数据集架构
所有领域的数据集均采用一致的架构,包含三个字段:
question:输入的问题或任务。reward_model:验证器或奖励模型信号的相关信息。teacher_ds:任务特定的专家轨迹,用于引导模型。
使用示例
可通过官方仓库提供的脚本,从原始专家轨迹构建掩码数据集。以下为下载并准备数学数据集的示例命令:
bash
下载原始数据
python scripts/data/download_from_hf.py --repo mit-han-lab/SMEPO --filename math.parquet --out-parquet data/raw/math_teacher.parquet
构建掩码数据集
bash scripts/data/build_data.sh
相关资源
- 论文地址:https://huggingface.co/papers/2605.25198
- 官方 GitHub 仓库:https://github.com/mit-han-lab/SMEPO




