Large Language Models in the Context of Higher Education
In this series of posts, I plan to explore some of the ways pre-trained language models (both large language models (LLMs) and increasingly, small language models (SLMs)) are transforming work in higher education, especially on the administrative side.
PLMs offer significant promise for higher ed, as they do in any field, but the unique context of education sometimes makes them difficult to deploy successfully. In general, any solution needs to navigate higher ed’s heightened focus transparency and fairness, work with limited resources, and have these essential characteristics:
Private & Secure: We need to ensure that student privacy is preserved and that no student data is shared with anyone outside the organization. FERPA protects this data, of course, and it is both our legal and ethical obligation to keep student data confidential. This almost always excludes closed-source, proprietary LLMs such as recent GPT models. (Even now, a few years after Gen AI has become commonplace, it can be a good idea to make sure no one is cutting and pasting student data — or any sensitive data — into Chat GPT.)
Free (or at least cheap): Most universities do not have a significant budget to invest in AI, especially given the current challenges facing the industry. In general, LLM solutions within higher ed will need to be open source, low-resource-intensive solutions; we’ll usually want free models that can be sustainably set up with at most a few days of employee labor.
Easy: In terms of both labor and sustainability, we will expect these solutions to require minimal effort. Once again, very few universities will have employee resources to devote to AI; usually, these will be add-on tasks, fit within an already busy staff schedule. Not all universities will have a knowledge-base or expertise in machine learning, either, and usually these solutions will be created by one or two individuals rather than a team of machine learning engineers. At least for now, we’re looking for any low-hanging fruit where LLMs can help us work more efficiently.
Fortunately, there are thousands of open-source models available on Huggingface, spanning a range of sizes and specialized for various common tasks. We can automate several time-consuming jobs quite easily and start integrating free LLMs into our processes, improving efficiency and creating new possibilities for student success. Our goal throughout these posts is to offer quick guides on how to build something that is easy, secure, and sustainable, and maybe even scalable.
Our first guide is on using LLMs to extract employee names from survey data.
