Web Scale Trustworthy Data for Machine Learning

如今，大型语言模型大多是基于英语的，并在大规模数据集上进行训练，其中爬行的网络文本数据是最大的元素。Mojeek是一个国际性的网络搜索引擎，拥有独立的技术栈，不断增长的搜索索引超过65亿个网页。Mojeek指数每天更新，数据可通过API获得。我们的API已被FAIR和其他公司用于开发包括聊天机器人在内的人工智能解决方案。在这个项目中，我们将为一个大型的第二阶段项目建立一个由研究人员和公司组成的联盟，该项目将使机器学习模型的训练能够轻松快速地访问大规模数据。我们将进行研发，开发Mojeek API，并研究聊天机器人和在线安全方面的行业用例。我们将开发能够提取可信数据源的方法和软件，从而加速开发负责任的人工智能解决方案，特别是中小企业。

Large Language Models are today mostly English-language-based and trained on large-scale datasets, for which crawled web text data is the biggest element. Mojeek is an international web search engine, with an independent technology stack, and a growing search index of over 6.5 billion webpages. The Mojeek index is updated daily and data is available through the API. Our API has been used by FAIR and others to develop AI solutions including chatbots. In this project we will build a consortium of researchers and companies, for a large Phase 2 project which will enable easy and rapid access to large-scale data for training of machine learning models. We will conduct R&D, develop the Mojeek API, and work on industry use cases in chatbots and online safety. We will develop methodologies and software that will enable the extraction of trusted sources of data, thus accelerating the development of responsible AI solutions, and notably by SMEs.