32 Association of Research Libraries Research Library Issues 298 — 2019 recommendations for workflow process improvement and digital preservation. These fellowships (part of a larger cohort throughout the library) provided practical, on-the-ground work experience for these individuals, enabling them to build on their education and prepare for their careers. They also provided NAL with our first data collection policy documents and protocols for curation at the Ag Data Commons. And they helped conduct an important assessment of federal responses to US public access policy.16 The partnership with UMD taught us valuable lessons about how to offer data curation services. We quickly realized that a combination of data set self-submission and expert data curation would be necessary to scale our services without a reduction in quality. We are now preparing for a distributed curation model.17 Expertise in metadata standards and tools is critical and requires different knowledge than traditional library metadata. We realized it was important to include our data curators as stakeholders in agile software development—they are critical links between end users and developers, and they are the power users of our software. One former student is now one of our professional data curators, and two others are working in related positions (including management) elsewhere. Data Science Services In 2017, master’s of information management students from University of Maryland joined us to launch the first iteration of an experiment in providing data science services. These students considered themselves data scientists. We tested the hypothesis that an information school such as UMD can adequately prepare students to apply their skills to real-world problems. We tested the idea that data scientists need not have domain knowledge to conduct their work. Finally, we tested the idea that staff, such as that in a library, could supervise data science work of individuals working on distinct projects and produce results within a year. In collaboration with a high school intern, one student built an impact tracking dashboard prototype for the Ag Data Commons—we later refactored it and it is still in use.18 The same