Introducing AI Sheets: No‑code dataset building and enrichment with open models

TL;DR

• AI Sheets is an open-source, no-code spreadsheet-like tool for building, transforming, and enriching datasets using AI models.
• It runs in the browser via a Hugging Face Space or can be downloaded and deployed locally from GitHub.
• You can use thousands of open models from the Hub through Inference Providers or local models (including gpt-oss from OpenAI) and iterate by editing cells to provide few-shot examples.

Context and background

Hugging Face released AI Sheets, a tool designed to let users interactively build and refine datasets using AI models without writing code. The interface resembles a familiar spreadsheet paradigm where each new column can be generated by composing a natural-language prompt that references existing columns. AI Sheets integrates tightly with the Hugging Face Hub and the open-source model ecosystem, offering access to models via Inference Providers or by running models locally. The tool is aimed at experimentation workflows: users start small with a few rows to refine prompts and behavior before scaling up to larger or costlier generation runs. AI Sheets can be tried immediately in a hosted Space or installed from source for local deployment.

What’s new

AI Sheets introduces a no-code environment for several dataset workflows:

• Creating datasets from scratch by describing the desired structure and content in natural language (an “auto-dataset” or “prompt-to-dataset” feature).
• Importing existing datasets and adding AI-generated columns for transformations, classifications, analyses, enrichment (including web searches when enabled), and synthetic data generation.
• Iterative prompt refinement directly in the spreadsheet: manual edits and validation actions are incorporated as few-shot examples for subsequent regeneration of column content.
• Model comparison and automated judging: add one column per model to compare outputs, and create a judge column that uses an LLM to evaluate or rank responses. The product is available to try without installation at the hosted demo: https://huggingface.co/spaces/aisheets/sheets and the code can be downloaded from the repository: https://github.com/huggingface/sheets.

Why it matters (impact for developers and enterprises)

AI Sheets lowers the barrier to rapid data-centric experimentation by providing a familiar spreadsheet interface combined with model-driven automation. Key impacts include:

• Faster iteration on prompts and data formats: teams can preview and refine outputs on a handful of rows before committing compute resources to large-scale generation.
• Flexible model access: the ability to use thousands of Hub models via Inference Providers or local models enables testing of open models and proprietary stacks in the same workflow. The release explicitly mentions support for models such as gpt-oss from OpenAI.
• Data augmentation and cleaning: tasks like removing punctuation, extracting ideas, categorizing text, or filling missing fields (e.g., zip codes) can be handled via simple prompt-driven columns.
• Reproducibility and portability: exporting datasets to the Hub produces a config file that can be reused for large-scale generation jobs and downstream applications. For enterprises and developers building data pipelines or preparing training data, AI Sheets offers a quick way to prototype data transformations, test multiple models against the same inputs, and capture human-in-the-loop corrections that improve automated outputs.

Technical details or Implementation

User interaction model

• Interface: AI Sheets uses a spreadsheet-like interface. Each dataset is shown as editable rows and columns where new columns are created by writing prompts. Prompts can reference existing columns using placeholder syntax (for example, {{text}}).
• Iteration loop: after generating a column, users can manually edit cells or “like” results. Those edits and likes are automatically included as few-shot examples the next time the column is regenerated or expanded.
• Column configuration: users can change the prompt, switch models or inference providers, toggle options such as “Search the web” for external lookups, and then regenerate the column. Model access and deployment
• Hosted demo: Try AI Sheets without installation at https://huggingface.co/spaces/aisheets/sheets.
• Local deployment: The source code is available at https://github.com/huggingface/sheets for users who want to run AI Sheets locally or self-host. For local use, Hugging Face recommends subscribing to PRO to obtain 20x monthly inference usage if you want to maximize inference capacity.
• Models: AI Sheets can call thousands of models available on the Hugging Face Hub via Inference Providers, or use local models. The announcement explicitly includes gpt-oss from OpenAI as an example of supported open models. Export and scale
• Exporting to the Hub: When you export a dataset to the Hugging Face Hub from AI Sheets, the tool generates a configuration file describing the prompts and the few-shot examples derived from edits/validations. That config can be reused to generate larger datasets with automated jobs or to reuse prompts in downstream applications. Practical examples and workflows
• Model comparison: import a dataset of prompts/questions and add one column per model, each with a prompt like: “Answer the following: {{prompt}}”. Optionally, add a judge column with a prompt that evaluates multiple model outputs.
• Data cleaning/transformation: to normalize text, add a column with a prompt such as “Remove extra punctuation marks from the following text: {{text}}” and regenerate after validating examples.
• Classification and analysis: add columns that categorize or extract key ideas with prompts like “Categorize the following text: {{text}}” or “Extract the most important ideas from the following: {{text}}”.
• Enrichment and web lookups: for tasks like finding missing zip codes, enable the “Search the web” option and use a prompt referencing an address column.
• Synthetic data generation: create a small prompt-to-dataset describing the structure (for example, a professional bio column), then generate dependent columns like realistic emails written by that persona. Below is a quick comparison of example tasks and the suggested column prompt style. | Task | Example column prompt |--- |--- |Answering prompts / model comparison | Answer the following: {{prompt}} |Cleaning text | Remove extra punctuation marks from the following text: {{text}} |Classification | Categorize the following text: {{text}} |Analysis / extraction | Extract the most important ideas from the following: {{text}} |Enrichment (web) | Find the zip code of the following address: {{address}} |

Key takeaways

• AI Sheets is a no-code, spreadsheet-style tool for dataset creation, transformation, enrichment, and model comparison using open models or local models.
• It supports iterative human-in-the-loop refinement by using manual edits and validations as few-shot examples.
• The tool can be tried on Hugging Face Spaces or deployed locally from the GitHub repository; exporting to the Hub generates reusable configs for scaling.
• Use cases include prompt tuning, model benchmarking, data cleaning, labeling, enrichment, and synthetic data generation.

References

• Try AI Sheets (hosted demo): https://huggingface.co/spaces/aisheets/sheets
• Source code and local deployment: https://github.com/huggingface/sheets
• Example dataset produced with AI Sheets: https://huggingface.co/datasets/dvilasuero/jsvibes-qwen-gpt-oss-judged
• Original announcement: https://huggingface.co/blog/aisheets