AI Sheets 是一款无需编写代码的工具，用于构建、转换和丰富数据集，支持（开放）AI 模型。该工具与 Hugging Face Hub 及开源 AI 生态系统深度集成。AI Sheets 采用类电子表格的简易用户界面，易于上手；其设计以快速实验为核心，鼓励用户先从小规模数据集入手，再逐步运行耗时或高成本的数据生成流水线。

AI Sheets is a no-code tool for building, transforming, and enriching datasets using (open) AI models. It’s tightly integrated with the Hub and the open-source AI ecosystem. AI Sheets uses an easy-to-learn user interface, similar to a spreadsheet. The tool is built around quick experimentation, starting with small datasets before running long/costly data generation pipelines.

AI Sheets 是一款无需编码的工具，用于借助（开源）AI 模型构建、转换和丰富数据集。它与 Hugging Face Hub 及开源 AI 生态系统深度集成。AI Sheets 采用易于上手的类电子表格用户界面，以快速实验为核心设计理念：先从小规模数据集入手，再逐步运行耗时或高成本的数据生成流程。在 AI Sheets 中，新列通过编写提示词（prompt）创建；您可以反复迭代，随时编辑或验证单元格内容，从而引导模型理解您的需求。更多细节后文详述！ AI Sheets 提供两种启动方式：导入现有数据，或从零开始生成数据集。数据加载完成后，您可通过添加列、编辑单元格以及重新生成内容等方式对数据进行精炼。入门时，您可选择用自然语言描述所需数据集并从零创建，或直接导入已有数据集。 适用场景：熟悉 AI Sheets 操作、头脑风暴、快速实验及构建测试数据集。此功能可视为“自动数据集”或“提示词生成数据集”——您只需描述需求，AI Sheets 即可自动生成完整的数据集结构与内容。该示例数据集仅含五行，但您可通过向下拖拽各列（包括图像列）来增加行数。您也可在任意单元格中手动输入条目，并通过拖拽填充其余单元格。 适用场景：绝大多数需对真实世界数据执行转换、分类、增强与分析的任务。推荐大多数用户优先采用此方式，因为导入真实数据比从零开始提供更强的控制力与灵活性。 专业提示：若所导入文件仅含少量数据，您可直接在电子表格中键入内容，手动补充更多条目。 数据加载完成（无论采用何种启动方式）后，您将在一个可编辑的电子表格界面中查看数据。以下是您需要了解的关键信息： 现在，您已创建了 AI 列，可进一步优化其输出结果并扩展数据规模。优化方式包括： 1. 通过手动编辑单元格或点击“点赞”提供反馈； 2. 调整列配置。 以上两种操作均需重新生成内容方能生效。底层机制为：这些经手动编辑或点赞的单元格将作为少样本（few-shot）示例，在您重新生成或新增列内单元格时被模型调用。 3. 调整列配置 修改提示词、切换模型或服务提供商、调整参数设置，随后重新生成以获得更优结果。 当您对新建数据集满意后，即可将其导出至 Hub！此举额外带来两项优势： （1）自动生成一份配置文件，可用于通过 HF Jobs 脚本批量生成更多数据； （2）复用其中的提示词（含您编辑与点赞过的少样本）于下游应用。 本节提供若干 AI Sheets 可构建的数据集示例，以激发您的下一个项目灵感。 若您希望在自己关心的不同提示词与数据上测试最新模型，AI Sheets 将是理想搭档。您只需导入一个数据集（或从零创建），然后为不同模型添加对应列。之后，既可人工检查结果，亦可新增一列，调用大语言模型（LLM）对各模型输出质量进行评估。 以下为一个实例：对比多个开源前沿模型在微型 Web 应用场景下的表现。AI Sheets 支持交互式查看结果，并可实时试用每个应用。此外，该数据集还包含若干由 LLM 驱动的列，用于评判与比较各应用的质量。 由前述会话导出的示例数据集地址： https://huggingface.co/datasets/dvilasuero/jsvibes-qwen-gpt-oss-judged AI Sheets 还可增强现有数据集，并助力文本数据集的快速分析与数据科学项目。另一典型用例是采用 LLM 作为“裁判”（judge）来评估模型输出质量。该方法适用于模型横向对比，或评估现有数据集质量（例如，基于 Hugging Face Hub 上的现有数据集开展模型微调）。 在首个示例中，我们将氛围测试（vibe testing）与裁判型 LLM 列相结合。裁判提示词如下： 若您有任何问题或建议，欢迎前往 Community 标签页反馈，或在 GitHub 上提交 issue。

AI Sheets is a no-code tool for building, transforming, and enriching datasets using (open) AI models. It’s tightly integrated with the Hub and the open-source AI ecosystem. AI Sheets uses an easy-to-learn user interface, similar to a spreadsheet. The tool is built around quick experimentation, starting with small datasets before running long/costly data generation pipelines. In AI Sheets, new columns are created by writing prompts, and you can iterate as many times as you need and edit the cells/validate cells to teach the model what you want. But more on this later! AI Sheets gives you two ways to start: import existing data or generate a dataset from scratch. Once your data is loaded, you can refine it by adding columns, editing cells, and regenerating content. To get started, you need create one from scratch describing it in natural language or import an existing dataset. Best for: Familiarizing with AI Sheets, brainstorming, rapid experiments, and creating test datasets. Think of this as an auto-dataset or prompt-to-dataset feature—you describe what you want, and AI Sheets creates the entire dataset structure and content for you. This dataset contains only five rows, but you can add more cells by dragging down on each column, including the image one! You can also write items in any of the cells and complete the others by dragging. Best for: Most use cases where you want to transform, classify, enrich, and analyze real-world data. This is recommended for most use cases, as importing real data gives you more control and flexibility than starting from scratch. Pro tip: If your file contains minimal data, you can manually add more entries by typing directly into the spreadsheet. Once your data is loaded (regardless of how you started), you'll see it in an editable spreadsheet interface. Here's what you need to know: Now that you have AI columns, you can improve their results and expand your data. You can improve results by providing feedback through manual edits and likes or by adjusting the column configuration. Both require regeneration to take effect. Under the hood, these manually edited and liked cells will be used as few-shot examples for generating the cells when you regenerate or add more cells in the column! 3. Adjust column configuration Change the prompt, switch models or providers, or modify settings, then regenerate to get better results. Once you're happy with your new dataset, export it to the Hub! This has the additional benefit of generating a config file you can reuse for (1) generating more data with HF jobs using this script, and (2) reusing the prompts for downstream applications, including the few shots from your edited and liked cells. This section provides examples of datasets you can build with AI Sheets to inspire your next project. AI Sheets is your perfect companion if you want to test the latest models on different prompts and data you care about. You just need to import a dataset (or create one from scratch) and then add different columns with the models you want to test. Then, you can either inspect the results manually or add a column to use LLMs to judge the quality of each model. Below is an example, comparing open frontier models for mini web apps. AI Sheets lets you see the interactive results and play with each app. Additionally, the dataset includes several columns using LLM to judge and compare the quality of the apps. Example dataset exported from a session like the one we just described: : https://huggingface.co/datasets/dvilasuero/jsvibes-qwen-gpt-oss-judged AI Sheets can also augment existing datasets and help you with quick data analysis and data science projects that involve analyzing text datasets. Another use case is evaluating the outputs of models using an LLM as a judge approach. This can be useful for comparing models or assessing the quality of an existing dataset, for example, fine-tuning a model on an existing dataset on the Hugging Face Hub. In the first example, we combined vibe testing with a judge LLM column. Here's the judge prompt: If you have questions or suggestions, let us know in the Community tab or by opening an issue on GitHub.