Grading Real Coursework with a Local Language Model

Grading programming coursework is challenging work, and sending student work to a commercial API raises privacy problems for a university. We built a grading system that runs entirely locally and tested it on 125 real submissions from six different laboratory exercises for which two instructors had already marked every submission so that variation between human graders was also represented. The system extracts a rubric deterministically, unzips and cleans each submission, grades against the rubric with a local 7B model, and splits any submission too large for the context window into per-file chunks that are graded separately and combined. We compared it against the instructors and against a frontier Claude model, run agentically rather than through our pipeline, as a benchmark. On the shorter exercises, the local 7B agreed with the instructors as closely as the frontier model agreed, both on the one laboratory with real instructor spread and on the near-flat ones where nearly every student received the maximum. The paired difference in mean absolute error is −0.06 points out of 20 (95% CI −0.27 to +0.15). On the large code-project labs, unpacking archives and chunking oversized work increased coverage from 63% to 93% of the 125 submissions, and the same small model graded work it could not previously process, achieving agreement within about two points of the instructor’s grades. Matching the model’s strictness to the instructor removed most of the leniency that language model graders are known for on the laboratory used to set it. Local grading is viable for real coursework, and a tiered design sends the rare giant project to a larger tier.