Research data for multilevel connectionist evaluation of morphological relation geometry in Kazakh legal language

This record contains the frozen benchmark data and analytical outputs supporting the article “Multilevel connectionist evaluation of morphological relation geometry: Word2Vec–FastText retrieval, transfer, and error locality in Kazakh legal language”. The study compares Word2Vec and FastText representations trained on the same Kazakh legal corpus. The deposit contains the conservative 200-pair morphological reference benchmark, 4,800 directional vector analogies across eight relation types, model predictions and rank-based evaluation outputs, relation-vector coherence and leave-one-out classification results, internal relation-centroid geometry, balanced and full cross-domain transfer datasets, dependency-aware robustness analyses, corrected statistical-validation outputs, rank-tail diagnostics, reproducible error-locality data, consolidated article tables, and final figures. Empty expert-review templates, internal pending-validation files, obsolete local paths, raw legal texts, and model binaries are not included. No analytical result reported in the article was recalculated or altered during preparation of the public package. No human-participant data or personal data are included.