Understand what the dataset measures
The original QM9 collection reports quantum-chemical properties for small organic molecules containing C, H, O, N and F, with up to nine heavy atoms. These are computed labels for a specified chemical domain, not measured drug activity or evidence of general performance on larger molecules. PyG provides a processed QM9 dataset with multiple targets, graph features and positions. Record the exact implementation revision, input files, exclusions and processed molecule count rather than assuming the original and processed inventories are identical.
qm9-collection · qm9-original-data · pyg-torch_geometric/datasets/qm9.py
Write a target contract
For the inspected PyG implementation, target index 0 is dipole moment μ in Debye (D). It is a scalar magnitude, not a three-component vector and not eV. Other targets have their own units: the loader documents energy conversions, so original-file units must not be assigned blindly to processed tensors. Record target name, index, source definition, units and any transformation. For dipole regression, report MAE in D after reversing target scaling; standardized training loss alone is not a comparable physical error.
pyg-torch_geometric/datasets/qm9.py
Match the input information
A bond-graph model and a model using supplied 3D coordinates answer different information-access questions. State whether geometry is available at prediction time and where it comes from. PyG provides positions, but their availability in a benchmark does not make experimental or calculated geometries freely available for future compounds. Keep conformer generation and selection inside the recorded protocol. Do not compare a 2D baseline with a 3D result as though both received identical information.
pyg-torch_geometric/datasets/qm9.py
Split first, fit preprocessing second
Choose random or chemical-family-separated evaluation according to the intended generalization question; neither split is universally correct. Resolve duplicates before allocating records. Freeze training, validation and test identifiers, fit learned normalization on training only, tune on validation, and evaluate the test set after selection. Report split files, seed, baseline, tuning budget, exclusions and repeated-run variability if actually measured. No scores are supplied by this guide, and a teaching Recipe is not evidence of benchmark-level performance.
pyg-torch_geometric/datasets/qm9.py
Trace the data and its rights
The official Figshare API identifies “Data for 133885 GDB-9 molecules,” item 1057646 version 2, as CC0. The separately inspected readme item is not a substitute for that data record. PyG code has a separate software license, and third-party mirrors may add conversions, exclusions or additional labels. Retain original item/DOI, version, file checksums, processing code and license evidence. Do not transfer a mirror’s CC BY label to the original dataset or treat a software license as data rights. QM9 conclusions remain limited by its chemical domain and chosen computational labels.
qm9-original-data · qm9-data · pyg-LICENSE
Related resources and reading
A QM9 Dipole GNN Teaching Baseline with PyTorch Geometric
Evaluating Molecular Property Prediction Models · Dataset Provenance and Licensing for Chemistry AI · Reproducibility for Chemistry AI Workflows
Sources and evidence boundary
Sources were inspected on 2026-10-08. Revision-pinned project documentation supports capability statements. Evaluation choices are editorial proposals. This article reports no executed workflow, measured performance, or experimental validation.