Seal-Tools

Seal-Tools是由苏州大学创建的一个新型工具学习数据集，旨在通过自指导方法生成大量可靠的工具实例，以精确控制数据生成过程。该数据集不仅包含大量工具，还包括展示工具实际应用的实例，特别设计了包含嵌套工具调用的复杂实例，以提高数据集的挑战性。Seal-Tools通过严格的格式控制和多维度评估指标，为LLMs的工具调用能力提供了一个全面的评估基准。数据集的应用领域主要集中在提升LLMs理解和使用外部工具的能力，以解决复杂任务。

Seal-Tools is a novel tool-learning dataset developed by Soochow University. It aims to generate a large quantity of reliable tool instances through self-instructing methods to enable precise control over the data generation process. This dataset includes not only a wide range of tools but also practical application examples showcasing how these tools are used; specifically, it designs complex instances involving nested tool calls to elevate the dataset's challenge level. Seal-Tools provides a comprehensive evaluation benchmark for the tool-calling capabilities of Large Language Models (LLMs) via strict format control and multi-dimensional evaluation metrics. Its primary application scenarios focus on improving LLMs' abilities to understand and employ external tools to solve complex tasks.