tinaxie/Uno-Curriculum

Uno-Curriculum数据集是一个用于训练分层委托路由器的语料库，该路由器是一个小型语言模型，能够将任务分解为子任务，并将每个子任务路由到特定的（工作模型，技能）对。数据集中的每一行都来自真实的公开HuggingFace数据集，问题和答案直接从源数据集中采样。每行数据经过三个阶段的处理流程（路由器探测→教师轨迹→噪声去除），以确保多轮对话轨迹的质量。数据集包含多种配置，适用于不同的使用场景，如主要训练、真实部署审计和子任务级路由分析。数据集主要为英文，采用Apache-2.0许可证。

The Uno-Curriculum dataset is a training corpus for a hierarchical-delegation router: a small language model that decomposes a task into subtasks and routes each subtask to a (worker model, skill) pair. Every row comes from a real public HuggingFace dataset, with questions and answers sampled verbatim from the source dataset. Each row undergoes a three-stage pipeline (router probe → teacher trajectory → noise removal) to ensure the quality of multi-turn trajectories. The dataset includes various configurations for different use cases, such as primary training, real-rollout audits, and subtask-level routing analysis. The dataset is primarily in English and is licensed under Apache-2.0.