py-irt: A Scalable Item Response Theory Library for Python.

The article presents py-irt, a Python library for fitting Bayesian item response theory (IRT) models, addressing the lack of scalable Python tools for large-scale IRT analysis. Built on the Pyro and PyTorch frameworks, py-irt leverages GPU acceleration and stochastic variational inference to efficiently estimate latent traits of subjects and items, supporting various IRT model types including one-, two-, and three-parameter logistic models. The package is designed for both practitioners and researchers, offering ease of use via command-line or Python interfaces and flexibility to define custom models. Experimental results demonstrate py-irt's superior scalability and comparable accuracy to established R packages like mirt, with applications shown in educational testing, machine learning evaluation, and natural language processing. py-irt is open-source, available on GitHub and PyPI, and aims to support future extensions such as multidimensional and graded response models.