Answers are grounded in this paper and cite their sources.
Machine-learning thermoelectric scores are difficult to compare when source provenance and split groups differ. TEBench organises Starrydata Experiment-labelled curves (30,530 samples, 1,648,315 points, 6,045 papers), a JARVIS-DFT extract (23,558 structures) and an audited bulk Bi2Te3 workbook (740 records, 151 DOIs) as separate domains. DFT and experiment are never pooled. Canonical composition keys, paper/family/formula grouping and an eight-paper own-data hold-out support leakage review. The original Bi2Te3 Seebeck random forest reaches test R2 = 0.935 but own-paper R2 = −1.21. A κ-specific physical screening rule retains 257,647 Starrydata points before composition parsing and split joining; the rerun gives test R2 = 0.045 (random forest) and 0.232 (gradient boosting). Validation-selected CPU gradient boosting and structure-derived JARVIS tabular features are evaluated beside preserved random-forest references. Composition-to-JID candidate matches are explicitly distinguished from unique structure identity. Extract hashes and transport-unit limitations are documented. These are proposed tasks and transparent baseline comparisons, not a released benchmark or a claim of state-of-the-art prediction.
No takes yet. Share an insight, caveat, or question.
Rapaka S. C. Bose (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: