Hugging Face has introduced the Data Measurements Tool (DMT), an interactive interface and open-source library designed to help dataset creators and users automatically calculate meaningful metrics for responsible data development. The tool aims to address the often-overlooked analysis of machine learning datasets by providing a no-code way to understand, build, curate, and compare datasets.
The DMT allows users to explore dataset basics such as descriptions and missing values, descriptive statistics like vocabulary size and word distribution, distributional statistics, and comparison statistics. It is built on Hugging Face's Datasets and Spaces Hubs, integrated with Streamlit.
This release comes amid calls for a paradigm shift in how the AI field approaches datasets, emphasizing the need for fine-grained requirements, curation for bias, and explicit values in construction. The DMT aims to lower the barrier for people from various disciplines to analyze datasets without requiring complex coding skills.
The tool currently supports iterative exploration of existing NLP datasets, with plans to support dataset creation from scratch soon. It draws inspiration from tools like Google's Know Your Data and IBM's Data Quality for AI, as well as research proposals such as Datasheets for Datasets.