A half-life number without its dose, population, analyte, and study context can be misleading. The useful contribution here is not just scale: the records preserve those distinctions, making literature-derived measurements easier to examine without pretending that unlike studies are directly comparable.
What Sets It Apart
- Breadth with provenance: 19,952 observations from 3,270 PubMed Central articles cover oral, intravenous, subcutaneous, and other routes. Each record is linked to its source article.
- Context beside the value: Records distinguish analyte, regimen, cohort, occasion, and summary statistic; reported values and uncertainty remain available as text alongside normalized hours.
- Chemical identity where resolvable: 12,434 records include paired SMILES and InChIKey identifiers. Missing identifiers are expected for biologics, mixtures, radioactivity, and unresolved analytes—not evidence that a record is invalid.
- Useful beyond compound lookup: The collection can support exploratory PK comparisons and research on literature extraction and normalization. It is not a ready-made train/test benchmark: the Hugging Face
trainsplit is a storage convention, not a validated partition.
Who It’s For—and the Trade-off
Great fit for researchers reviewing reported half-lives, building extraction workflows, or generating hypotheses from published PK data. Look elsewhere for prescribing guidance, dose selection, or a clean benchmark requiring standardized experimental conditions: methods, populations, sampling, and terminal-phase definitions vary, and the dataset explicitly warns against treating records as automatically comparable.
