ML_UV_models: v3 artifact bundle for "When Do Simple Models Win? Machine Learning Architectures for UV Absorption Prediction"

Companion artifacts for the manuscript "When Do Simple Models Win? Machine Learning Architectures for UV Absorption Prediction" (J. Chem. Inf. Model., manuscript ci-2026-009433). Source code, training scripts, and analysis notebooks: github.com/Umesh1608/ML_UV_models Bundle contents data/ v3 cleaned primary dataset (Greenman-Song dedup) data/UV_canonical_v3_dedup.csv -- 18,415 (solute, solvent) pairs fold_indices/ per-fold stratified CV index file (5-fold) weights/ trained model weights per fold: rf/ rf_tuned_v3_fold{0..4}_model.joblib ~4.7 GB xgboost/ xgboost_v3_fold{0..4}_model.joblib ~7 MB chemprop/ chemprop_v3_fold{0..4}_best.ckpt (default config) ~50 MB bigru/ bigru_tuned_v3_fold{0..4}_best.keras ~115 MB chemberta/ chemberta_v3_fold{0..4}_best.pt sizes vary predictions/ per-fold test-set predictions + ground truth + indices into the CSV metrics/ per-fold RMSE/MAE/R^2, CV aggregates, and the exact hyperparameter configs used What this enables Reviewers can (a) reproduce the headline RMSE numbers reported in Table 2 of the manuscript without retraining (using the saved per-fold predictions); (b) load any saved checkpoint and run inference on new molecules; and (c) verify the chromophore-overlap analysis, the non-local-subset breakdown, and the wetlab scaffold audit. Chemprop tuned The Chemprop tuned configuration (5.6M params; reported as a separate row in Table 2) is reproducible from default-config training scripts plus the saved hyperparameters in metrics/chemprop_tuned_v3_config.json. The default Chemprop checkpoints in weights/chemprop/ reproduce the bolded primary Table 2 row. License MIT License, matching the GitHub repository.