Hugging Face AI Sheets Release

Hugging Face AI Sheets is an open-source, no-code tool designed for building, enriching, and transforming datasets using AI models. It enables users to leverage thousands of open models from the Hugging Face Hub via Inference Providers or local deployments, including models such as gpt-oss from OpenAI.

No-Code Dataset Engineering with AI Sheets

AI Sheets provides a spreadsheet-like user interface that allows users to create new data columns by writing prompts. This approach facilitates rapid experimentation, allowing users to start with small datasets and iterate on prompts before scaling to larger data generation pipelines.

Core Capabilities

Users can employ AI Sheets for a variety of data processing and generation tasks:

  • Model Comparison and "Vibe Testing": Users can import prompts and create multiple columns—one for each model—to compare outputs side-by-side. This can be extended by adding a "judge" column where another LLM evaluates the responses.
  • Prompt Optimization: The tool allows for efficient prompt fine-tuning. By editing cells or marking them with a "thumbs-up," users provide direct feedback that the tool automatically incorporates as few-shot examples for subsequent generations.
  • Dataset Transformation and Classification: AI Sheets can be used to clean text (e.g., removing punctuation), categorize content, or extract key ideas from a dataset.
  • Data Enrichment: By enabling the "Search the web" option, users can use models to find missing information, such as zip codes for specific addresses.
  • Synthetic Data Generation: Users can generate entirely new datasets from scratch by describing the desired structure in natural language or by chaining prompts (e.g., generating a professional bio and then using that bio to write a realistic email).

Workflow and Technical Implementation

AI Sheets supports two primary entry points for data: importing existing files (XLS, TSV, CSV, or Parquet) up to 1,000 rows, or generating a dataset from scratch using a natural language description.

Refinement and Expansion

Once data is loaded, users can refine the output through several mechanisms:

  • Manual Feedback: Editing a cell or liking a result serves as a few-shot example for the model during regeneration or when adding new rows.
  • Column Configuration: Users can modify the prompt, switch models, or change inference providers to optimize performance.
  • Dynamic Expansion: Users can drag down from the last cell in a column to instantly generate additional rows.

Hub Integration and Scaling

Datasets created in AI Sheets can be exported to the Hugging Face Hub. This export generates a configuration file (config.yml) that preserves the prompts and few-shot examples. This configuration can then be used to scale data generation using HF Jobs and a provided extension script to process larger volumes of data beyond the initial spreadsheet limit.

Deployment and Access

AI Sheets can be accessed as a free tool via Hugging Face Spaces or deployed locally via its GitHub repository. For local users, Hugging Face recommends a PRO subscription to increase monthly inference usage.

Sources