LLM360/guru-RL-v1.5

Guru (v1.5) 是 Guru-92k 的一个子集，专注于数学、代码和科学数据。该数据集经过特别筛选，去除了用于中期训练和SFT阶段的数据以及下游评估基准中的重复数据，用于开发 K2 Think V2。数据集以 Parquet 格式存储，每个样本包含多个字段，如数据来源、提示、能力、聊天模板应用标志、特定模型的通过率、额外信息和奖励模型。数据集涵盖数学、代码和科学三个领域，总样本量为 108.8K。数学领域包括竞赛问题和符号推理，代码领域包括编程挑战和算法设计，科学领域包括大学/博士级别的物理、化学和生物学问题。

Guru (v1.5) is a focused subset of Guru-92k using only Math, Code, and Science data. This mid-version update was specially filtered according to the capabilities of K2-V2, deduplicated from the data used for mid-training and SFT stages as well as from any downstream evaluation benchmarks, for its post-training in development of K2 Think V2. The dataset is stored in Parquet format for efficient access and processing. Each sample contains fields such as data_source, prompt, ability, apply_chat_template, pass rates with specific models, extra_info, and reward_model. The dataset covers Math, Code, and Science domains, with a total of 108.8K samples. Math domain includes competition problems and symbolic reasoning, Code domain includes programming challenges and algorithm design, and Science domain includes university/PhD-level physics, chemistry, and biology problems.