Hugging Face has released AI Sheet, an open source application designed to build, transform and enrich datasets using AI models via spreadsheet-like interfaces. The tool is available in both hubs and local deployments, allowing users to experiment with thousands of open models, including Openai's GPT-Oss, without the need for code.
The interface is similar to a traditional spreadsheet, but instead of manual expressions, new columns can be generated at the prompt. For example, users can clean text, categorize entries, enhance datasets that lack detail, or generate synthetic lines by describing the desired output in natural language. Cells can be directly edited or verified, which can lead to models in subsequent generations.
The AI sheet supports two entry points. Generating a dataset involves writing its structure in a simple language or importing an existing dataset in CSV, TSV, XLS, or parquet format. The first option is good for prototyping or synthetic data generation, but importing the actual data will result in large-scale transformation and enrichment tasks. The embracing face emphasizes that the experiment begins with a small sample before scaling into a larger pipeline.
This tool also provides a mechanism for model comparison. Users can create multiple output columns, each with a different model, or add another column where another LLM acts as a judge to evaluate the results. In one published example, the researchers compared the outputs of the interactive mini-web apps QWEN3-Coder and GPT-oss with automatically generated ratings.
Some early adopters highlight both potential and limitations.
It has an LLM, but it's very slow. Why would someone prefer it over OpenRefine?
Concerns about data privacy have also emerged:
Is it possible to self-host this app? It sounds interesting, but there is no way to upload business data to a remote server. sorry.
In response, machine learning engineer Daniel Vila Suero, who hugs Face, confirmed that self-horst is supported.
Yes, you can self-host with Docker. See how you deploy to your hub space.
Once your dataset is refined, you can export it directly to a hugging facehub. The process also generates reusable configuration files that can be expanded using facial jobs that embrace the same pipeline, or by integrating them into downstream workflows.
AI sheets are available on the Hug Face Hub for free use without installation. The code can be deployed locally via GitHub.
