文章主题：moss-moon-003, moss-moon-003-sft-plugin, moss-moon-003-pm
IT之家 4 月 21 日消息，复旦大学自然语言处理实验室开发的新版 MOSS 模型今日正式上线，成为国内首个插件增强的开源对话语言模型。
目前，MOSS 模型已上线开源，相关代码、数据、模型参数已在 Github 和 Hugging Face 等平台开放，供科研人员下载。
据介绍，MOSS 是一个支持中英双语和多种插件的开源对话语言模型，moss-moon 系列模型具有 160 亿参数，在 FP16 精度下可在单张 A100 / A800 或两张 3090 显卡运行，在 INT4/8 精度下可在单张 3090 显卡运行。MOSS 基座语言模型在约七千亿中英文以及代码单词上预训练得到，后续经过对话指令微调、插件增强学习和人类偏好训练具备多轮对话能力及使用多种插件的能力。
MOSS 来自复旦大学自然语言处理实验室的邱锡鹏教授团队，名字与《流浪地球》电影中的 AI 同名，已发布至公开平台（https://moss.fastnlp.top/），邀请公众参与内测。
IT之家查看 MOSS 的 GitHub 页面发现，该项目所含代码采用 Apache 2.0 协议，数据采用 CC BY-NC 4.0 协议，模型权重采用 GNU AGPL 3.0 协议。如需将该项目所含模型用于商业用途或公开部署，需要签署文件并发送至 robot@fudan.edu.cn 取得授权，商用情况仅用于记录，不会收取任何费用。
MOSS 用例：
Moss Moon Base: MOSS-003 is a high-quality base model trained on self-supervised pre-training data, which contains approximately 700 billion words and 6.67 x 10^22 floating-point operations.
moss-moon-003-sft: 基座模型在约 110 万多轮对话数据上微调得到，具有指令遵循能力、多轮对话能力、规避有害请求能力。
🌟 Moss Moon 003 SFT Plugin: 🌟🚀 The base model was fine-tuned on over 110,000 rounds of dialogue data and approximately 30,000 plugin enhancements. It also has the ability to use search engines, generate text from images, perform calculations, and solve equations.💡 Moss Moon 003 SFT is built upon the foundation of Moss Moon 003, adding four additional features: search engines, image-to-text generation, calculator, and equation-solving capabilities.
🚀 Moss Moon 003 PM 🚀🚀 Moss Moon 003 PM 🚀🚀 Moss Moon 003 PM 🚀🚀 Moss Moon 003 PM 🚀🚀 Moss Moon 003 PM 🚀🚀 Moss Moon 003 PM 🚀🚀 Moss Moon 003 PM 🚀🚀 Moss Moon 003 PM 🚀🚀 Moss Moon 003 PM 🚀🚀 Moss Moon 003 PM 🚀🚀 Moss Moon 003 PM 🚀🚀 Moss Moon 003 PM 🚀🚀 Moss Moon 003 PM 🚀🚀 Moss Moon 003 PM 🚀🚀 Moss Moon 003 PM 🚀🚀 Moss Moon 003 PM 🚀🚀 Moss Moon 003 PM 🚀🚀 Moss Moon 003 PM 🚀🚀 Moss Moon 003 PM 🚀🚀 Moss Moon 003 PM 🚀🚀 Moss Moon 003 PM 🚀🚀 Moss Moon 003 PM 🚀🚀 Moss Moon 003 PM 🚀🚀 Moss Moon 003 PM 🚀🚀 Moss Moon 003 PM 🚀🚀 Moss Moon 003 PM 🚀
🚀 Moss Moon 003 is the final model trained on the basis of Moss Moon 003-SFT, which has better factualness and security as well as more stable response quality. It will be released soon.
🚀 Moss Moon 003 Plugin: A Stronger Intent Understanding and Plugin Usage Capable Model Trained on Moss Moon 003 SFT Plugin’s Foundation 🌟🚀 Moss Moon 003 Plugin is a powerful tool that has been trained using the Moss Moon 003 SFT Plugin as its foundation. This advanced model has undergone extensive training to enhance its ability to understand user intent and make better use of plugins.🚀 The Moss Moon 003 PM, which stands for Moss Moon 003 Pre-Training Model, plays a crucial role in the training process. It helps to provide a strong starting point for the model, allowing it to learn from vast amounts of data and improve its performance over time.🚀 With this powerful combination of Moss Moon 003 SFT Plugin and Moss Moon 003 PM, the Moss Moon 003 Plugin has become an even more capable tool. It is now able to provide users with a seamless experience when using plugins, making it easier than ever before to achieve their goals.🚀 We are excited to announce that this advanced model will soon be open source. This means that anyone can access and use the Moss Moon 003 Plugin, regardless of their technical expertise or background.🚀 So if you’re looking for a powerful tool that can help you achieve your goals more efficiently, look no further than the Moss Moon 003 Plugin. With its advanced capabilities and open-source nature, it’s sure to be a valuable asset in any workflow.
数据# Moss-002 SFT DataMoss-002 SFT Data is a comprehensive dataset that covers three key aspects of conversation quality: usefulness, loyalty, and harmlessness. It includes approximately 570,000 English conversations generated by text-davinci-003 and 590,000 Chinese conversations.This dataset provides valuable insights into the effectiveness of different conversational strategies and can be used to improve the overall quality of conversations in various contexts. Whether you’re working on a chatbot or developing a virtual assistant, Moss-002 SFT Data is an essential resource for anyone looking to enhance their conversation capabilities.# Key AspectsThe Moss-002 SFT Data dataset covers three key aspects of conversation quality: usefulness, loyalty, and harmlessness. These aspects are critical in determining the overall effectiveness of a conversational system.Usefulness refers to the ability of a conversation to provide useful information or assistance to users. A conversational system that is useful will be more likely to engage users and generate positive outcomes.Loyalty refers to the degree to which users feel connected to and committed to a conversational system. A loyal user is more likely to continue using the system and recommend it to others.Harmlessness refers to the potential negative impact of a conversation on users or society as a whole. A conversational system that is harmless will be more likely to be accepted by users and avoid causing harm.# English ConversationsThe Moss-002 SFT Data dataset includes approximately 570,000 English conversations generated by text-davinci-003. These conversations cover a wide range of topics and can provide valuable insights into the effectiveness of different conversational strategies.English conversations are an important aspect of the Moss-002 SFT Data dataset because they represent a diverse range of perspectives and experiences. By analyzing English conversations, researchers can gain a better understanding of how different cultures and communities interact with each other.# Chinese ConversationsThe Moss-002 SFT Data dataset also includes approximately 590,000 Chinese conversations generated by text-davinci-003. These conversations cover a wide range of topics and provide valuable insights into the effectiveness of different conversational strategies in Chinese culture.Chinese conversations are an important aspect of the Moss-002 SFT Data dataset because they represent a unique linguistic and cultural context. By analyzing Chinese conversations, researchers can gain a better understanding of how different cultures and communities interact with each other in their own language.# ConclusionIn conclusion, the Moss-002 SFT Data dataset is an essential resource for anyone looking to enhance their conversation capabilities. It covers three key aspects of conversation quality: usefulness, loyalty, and harmlessness, and includes approximately 570,000 English conversations generated by text-davinci-003 and 590,000 Chinese conversations.By analyzing these conversations, researchers can gain valuable insights into the effectiveness of different conversational strategies and improve the overall quality of conversations in various contexts. Whether you’re working on a chatbot or developing a virtual assistant, Moss-002 SFT Data is an essential resource for anyone looking to enhance their conversation capabilities.
moss-003-sft-data: moss-moon-003-sft 所使用的多轮对话数据，基于 MOSS-002 内测阶段采集的约 10 万用户输入数据和 gpt-3.5-turbo 构造而成，相比 moss-002-sft-data，moss-003-sft-data 更加符合真实用户意图分布，包含更细粒度的有用性类别标记、更广泛的无害性数据和更长对话轮数，约含 110 万条对话数据。目前仅开源少量示例数据，完整数据将在近期开源。
moss-003-sft-plugin-data: moss-moon-003-sft-plugin 所使用的插件增强的多轮对话数据，包含支持搜索引擎、文生图、计算器、解方程等四个插件在内的约 30 万条多轮对话数据。目前仅开源少量示例数据，完整数据将在近期开源。
moss-003-pm-data: moss-moon-003-pm 所使用的偏好数据，包含在约 18 万额外对话上下文数据及使用 moss-moon-003-sft 所产生的回复数据上构造得到的偏好对比数据，将在近期开源。
MOSS 的 GitHub 页面：点此查看
AI时代，掌握AI大模型第一手资讯！AI时代不落人后！
免费ChatGPT问答，办公、写作、生活好得力助手！
扫码右边公众号，驾驭AI生产力！
