AvitoTech/ru_bfcl
收藏资源简介:
Berkeley Function Calling Leaderboard(BFCL)数据集是一个用于评估不同大型语言模型(LLM)在函数调用(或工具调用)能力的数据集,基于实际使用案例构建,覆盖多种类别和语言。数据集分为三个主要版本:BFCL V1、V2 Live和V3。V1版本包括Python类别(简单函数、多函数、并行函数、并行多函数)和非Python类别(聊天能力、函数相关性检测、REST API、SQL、Java、JavaScript),旨在评估模型在常见API调用场景中的表现。V2 Live版本使用真实世界数据,更侧重于多函数场景和相关性/不相关性检测,包含简单、多函数、并行、并行多函数以及相关性检测子类别。V3版本引入了多步和多轮函数调用场景,包括基础多轮、增强多轮(缺失参数、缺失函数、长上下文、复合)等类别,以测试模型在复杂交互中的能力。数据集以JSON文件组织,支持AST和可执行评估,用于衡量LLM在函数调用任务中的泛化性和实用性。
The Berkeley Function Calling Leaderboard (BFCL) dataset is a live leaderboard to evaluate the ability of different LLMs to call functions (also referred to as tools). It is built from real-world use cases, spanning diverse categories and multiple languages. The dataset is organized into three major releases: BFCL V1, V2 Live, and V3. V1 includes Python categories (Simple Function, Multiple Function, Parallel Function, Parallel Multiple Function) and non-Python categories (Chatting Capability, Function Relevance Detection, REST API, SQL, Java, JavaScript) to assess model performance in common API call scenarios. V2 Live uses real-world data with a focus on multiple function scenarios and relevance/irrelevance detection, including simple, multiple, parallel, parallel multiple, and relevance detection subcategories. V3 introduces multi-step and multi-turn function calling scenarios, covering Base Multi-Turn, Augmented Multi-Turn (Missing Parameters, Missing Functions, Long-Context, Composite), etc., to test model capabilities in complex interactions. The dataset is structured in JSON files and supports both AST and executable evaluations, aiming to measure the generalization and practicality of LLMs in function calling tasks.




