MLentory: NFDI4DS registry for machine learning models and related artifacts

Data Science and Artificial Intelligence, particularly machine/deep learning (ML) models, have become increasingly common in research offering powerful and accurate solutions for predictions, classifications, and clustering. However, the rapid explosion of ML models and different options to share them, make it difficult for researchers to find relevant models - whether similar to their own or ready to be used in their own research. As a result, researchers often need help navigating through such a large and heterogeneous collection of ML models. MLentory, part of the NFDI4DS portfolio, addresses this challenge by providing a comprehensive and harmonized solution for ML model discovery. MLentory [1] is a registry built on an Extraction, Transformation, and Loading (ETL) pipeline, complemented with a back-end providing a REST API and a front-end, facilitating seamless access to developers, end-users, and machines. This is enabled through Webby FAIR Digital Objects (FDOs) [2] and metadata based on schema.org [3]. It aggregates, harmonizes, and stores metadata from various ML model repositories such as Hugging Face [8] and OpenML [9], with plans to integrate more, including AI4Life [10] and the DOME registry [11]. The metadata is processed using FAIR4ML and Croissant ML, structured into Webby FDOs, and indexed for efficient natural language-based conversational search. Webby FDOs are an implementation of FDOs using the common web and Linked Open Data technologies, in particular, Research Object Crate (RO-crate) [4] to pack research artifacts with their metadata and FAIR Signposting [5] for guiding machines to the key elements such as metadata, citations, and types. MLentory uses FAIR4ML [6] to describe ML models and Croissant ML [7] for dataset descriptions, both extending schema.org for enhanced interoperability. MLentory ETL pipeline, implemented using Python scripts, extracts metadata from various sources, transforms it into a standardized format and loads it into a PostgreSQL database for historical version tracking. It uses a Virtuoso database for RDF-based knowledge representation and an Elasticsearch module for efficient data indexing. Each stage of the pipeline operates independently within its own Docker container. The back-end, built with FastAPI v0.115.11 [8], serves as a query engine, enabling developers to programmatically retrieve information from the different data stores in MLentory. The natural language-based search leverages Elasticsearch for initial retrieval, relying on a self-hosted Large Language Model (LLM) powered by Ollama [9] to refine search results through Retrieval Augmented Generation (RAG). Finally, the front-end module, developed using Vue3.js v3.15.1 [10], provides a user-friendly interface that allows users to explore models using natural language, filter them based on multiple criteria, e.g., ML task or license, and access their metadata version history. MLentory will be officially launched during the second quarter of 2025. Its development will continue to improve retrieval, implement a recommendation system, and include more sources.