Code, data sample and results for "How much does distribution shift cost Arabic sentiment models on Saudi app store reviews? A matched-test benchmark on GLARE"

Code, a 52,689-review sample of the GLARE corpus of Saudi Google Play reviews, per-fold results, out-of-fold predictions, and logs for a benchmark of nine CPU-based Arabic sentiment models under five evaluation protocols: random, app-disjoint, leave-one-category-out, temporal, and games versus non-games. Each protocol is run with five seeds. The deposit includes the matched-test gap analysis, which compares shifted and randomly split models on the same test reviews. The review text is derived from GLARE (AlGhamdi et al., doi:10.5281/zenodo.6457824, CC BY 4.0). See README.md for contents and reproduction steps.