A Synthetic Benchmark and Real Reference Instance for Educational-Clinical Internship Scheduling with Reciprocal Supervision

This dataset provides a reproducible benchmark for internship scheduling in university psychology clinics, a scheduling problem whose distinctive feature is reciprocal supervision: one student conducts a clinical session while a peer observes, and the roles are later exchanged, with three-way (triadic) observation cycles arising when no directly compatible pair exists. This relational coupling between two trainees, rather than between an actor and a resource, is absent from existing public scheduling and timetabling benchmarks. The dataset contains two components. The first is an anonymized real reference instance migrated from a university psychology clinic (141 students, 12 supervisors, 6 rooms, 4 sequential internship stages, 65 weekly time periods), carrying only synthetic identifier codes and no personally identifiable information. The second is a synthetic suite of 60 instances generated by a parametrized generator calibrated to the real instance, organized into six families that independently vary scale, availability density, temporal clustering, curriculum load, period capacity, and triadic stress (a ring-structured availability pattern that makes binary supervision pairs scarce while keeping triadic cycles feasible). Every instance is released in three interoperable formats generated from the same source: a relational SQLite database, per-table CSV files, and a JSON metadata record, together with a per-instance report. The package also includes the relational schema, the parametrized generator, the socio-academic attribute sampler, and the suite-orchestration script, so that the entire suite can be regenerated bit-for-bit from a single seeded configuration. Each instance is characterized by structural metrics computed directly from its data, independently of any solver: availability density, demand-to-slot ratio, and the pairing feasibility rate, defined as the edge density of a stage-specific temporal-compatibility graph (the fraction of same-stage student pairs whose availability windows intersect), together with triangle counts and triadic-necessity measures for studying triadic supervision. These metrics support research on instance difficulty, operational research and mathematical programming, constraint programming, metaheuristics, and difficulty-aware machine learning. Each student also carries socio-academic attributes (age, active semester, preferred shift, weekly work hours, commuting time, undergraduate-research participation, and prior course and stage failures) that are fully synthetic but calibrated to the marginals and principal correlations of the real cohort; no real student record is reproduced, and a per-student label table is reserved for supervised learning.