Continuous Integration for Reproducible Shared Tasks with TIRA.io.

A major obstacle to the long-term impact of most shared tasks is their lack of reproducibility. Often only the test collections and the papers of the organizers and participants are published. Third parties who want to independently evaluate the state of the art for a task on other data must re-implement the participants' software. The tools developed to collect software from participants in shared tasks only partially verify its reliability at the time of submission, much less long-term, and do not enable third parties to reuse it later. We have overhauled the TIRA Integrated Research Architecture to address all of these issues. The new version simplifies task setup for organizers and software submission for participants, scales from a local computer to the cloud, supports ondemand resource allocation up to parallel CPU and GPU processing, and enables export for local reproduction with just a few lines of code. This is achieved by implementing the TIRA protocol with an industrystandard continuous integration and deployment (CI/CD) pipeline using Git, Docker, and Kubernetes.