MakamGraph: Code, Evaluation Splits, and Analysis Resources

This archive provides the source code, fixed evaluation splits, aggregate predictions, and statistical analysis resources accompanying the study “MakamGraph: Evaluating Relational Representations for Symbolic Turkish Makam Classification.”

MakamGraph represents symbolic compositions as directed, duration-weighted pitch-transition graphs. The archive also includes a graph-aware Transformer that injects composition-specific same-pitch, forward-transition, reverse-transition, interval-proximity, and two-hop relational biases into self-attention.

The experiments compare low-order statistical representations, matched and graph-aware Transformers, piece- and window-level graph-native models, transition controls, composer-disjoint evaluation, duplicate audits, input-length-matched comparisons, and a validated transition-bias ablation. The controlled evaluation subset contains 711 compositions from nine Turkish makam classes, with 79 compositions per class.

The original SymbTr corpus is not redistributed. Users must obtain SymbTr separately under its original terms. Exact composition identities, original stratified fold assignments, composer-disjoint fold assignments, portable corpus-relative paths, aggregate results, and the package versions used in the final experiments are included to support reproducibility.