PUBLIC DATA · RESEARCH TOOLS
Before comparing model scores, I need to check whether the same molecule appears more than once and how the training and evaluation data are separated. Here are eight previews I prepared from public molecular datasets, with structures, identifiers, sources and split information.
Each preview contains 150 or 200 rows for inspecting the columns and workflow. They are not the full datasets or performance benchmarks; the original measurements and calculations come from the cited researchers and providers.
What the tables add
SMILES represents a molecular structure as text; InChIKey is a structure identifier. A scaffold describes a shared molecular framework. Splitting by scaffold lets me examine how closely related structures are distributed between training and evaluation data.
The processing differs by dataset. Each linked data card explains the original sources, changes, split columns and terms of use. The licenses listed below apply to the public previews; check the source citations and notices as well.
Eight molecular-data previews
Tools for reading and comparing records
TrialTrace is a locally runnable tool for exploring version dates and changed fields in frozen records from 400 clinical trials. Uibang Seongdo preserves source identifiers from Korean traditional-knowledge records so readers can explore same-named formulas separately and compare their ingredients.
I also published example code for AqSolDB and FreeSolv ↗. Use the individual Hugging Face cards for the other previews.
Sources and use
These previews do not imply affiliation with or endorsement by the original authors or providers. Cite the source data as well as the preview. In particular, hERG screening signals are not clinical-safety findings, and the small previews should not be used to rank model performance.