TDG-Bench Leaderboard

Rankings — sorted by accuracy

#	Model	Accuracy	Macro F1

Accuracy & F1 comparison

Evaluation protocol

TaskSemi-supervised node classification

Classes8 semantic types — 1,032 annotated nodes

Data split10 / 10 / 80 — train / val / test, stratified

Seeds42, 123, 456, 789, 1000 — mean ± std reported

Node initializationall-MiniLM-L6-v2 (sentence-transformers)

OptimizerAdam, lr=0.01, weight_decay=5×10−&sup4;

RegularizationDropout 0.5, early stopping (patience=20)

Hidden dimsTuned over {64, 128, 256}