Manuscript-Scoped Cleaned Analysis Data for Multitask Learning Improves Aquatic and Soil Ecotoxicity Prediction

A frozen, manuscript-scoped cleaned-data snapshot supporting the no-metal W00 aquatic and M00 soil ecotoxicity analyses. The deposit contains exact model-ready records, registered outer split assignments, manuscript-scoped ECOTOX identifiers and aggregation/QC provenance, a column-level data dictionary, integrity manifests, and compact aggregate numerical audit tables. W00 contains 219,818 records across 100 task heads on the pTox = -log10(mol/L) scale; M00 contains 11,532 records across 16 task heads on the neg_log10_mol_kg = -log10(mol/kg) scale. The full cleaned relational database, M10, unrelated and historical experiments, rejected-record inventories, external structure dictionaries, row-level predictions, fitted preprocessing objects, model checkpoints, caches, and future-project variables are not included. CC BY 4.0 applies only to the authors' original selection, organization, documentation, and aggregate audit tables. Underlying U.S. EPA ECOTOX records and third-party chemical-structure information retain their source terms and are not relicensed.