DroidCall
收藏资源简介:
DroidCall是由北京邮电大学创建的第一个用于大语言模型(LLM)驱动的Android意图调用的训练和测试数据集。该数据集包含10,000个样本,涵盖了广泛的系统功能和第三方应用。通过预定义的函数封装意图调用过程,DroidCall采用高度灵活和可重用的数据生成管道,自动生成数据并进行格式验证和去重,确保数据的准确性、相关性和多样性。DroidCall旨在通过微调小型语言模型(如Qwen2.5-3B和Gemma22B)来提高Android意图调用的准确性,解决现有模型在准确意图调用方面的不足。
DroidCall, developed by Beijing University of Posts and Telecommunications, is the first training and testing dataset for large language model (LLM)-driven Android intent invocation. This dataset contains 10,000 samples covering a broad spectrum of system functions and third-party applications. By encapsulating the intent invocation process through predefined functions, DroidCall adopts a highly flexible and reusable data generation pipeline to automatically generate data, perform format validation and deduplication, thereby ensuring the accuracy, relevance and diversity of the dataset. DroidCall aims to improve the accuracy of Android intent invocation by fine-tuning small-scale language models such as Qwen2.5-3B and Gemma22B, addressing the shortcomings of existing models in accurate intent invocation.

- 1DroidCall: A Dataset for LLM-powered Android Intent Invocation北京邮电大学 · 2024年



