Kukedlc/web-design-diamond
收藏资源简介:
Web Design Diamond SFT数据集是一个用于训练小型语言模型(2B-8B参数)的指令微调数据集,专门用于根据自然语言指令生成高质量、自包含的HTML页面。该数据集具有以下核心特点:1)顺序推理机制,模型在实施前进行思考规划,避免信息泄露;2)真实工具使用,模型可以浏览和获取真实品牌设计系统(如Stripe、Linear、Tesla等)的实际设计数据;3)多轮对话支持,30%的示例包含修改请求,需要更新完整页面;4)双语内容,50%西班牙语和50%英语,西班牙语包含正确的重音和ñ字符;5)自包含输出,生成单个index.html文件,包含4-5个部分,无需构建即可运行。数据集包含500个示例(250个西班牙语,250个英语),其中150个为多轮对话,350个示例使用工具调用。数据格式为JSONL,包含推理内容、工具调用和完整的对话历史,适合直接用于监督微调训练。
The Web Design Diamond SFT Dataset is an instruction tuning dataset for training small language models with parameters ranging from 2B to 8B, specifically designed to generate high-quality, self-contained HTML pages based on natural language instructions. The core features of this dataset are as follows: 1) Sequential reasoning mechanism: The model conducts thinking and planning prior to implementation to avoid information leakage; 2) Real-world tool usage: The model can browse and access actual design data from real brand design systems such as Stripe, Linear, Tesla, etc.; 3) Multi-turn dialogue support: 30% of the examples include modification requests that require updating the complete web page; 4) Bilingual content: 50% Spanish and 50% English, with Spanish text containing correct accents and the ñ character; 5) Self-contained output: Generates a single index.html file consisting of 4-5 sections, which can run directly without additional building steps. The dataset contains 500 total examples (250 in Spanish and 250 in English), among which 150 are multi-turn dialogue instances and 350 examples use tool calls. The data is formatted as JSONL, including reasoning content, tool invocation records and complete conversation histories, making it suitable for direct supervised fine-tuning training.




