lsieben/climb-results

CLIMB evaluation results

Per-run evaluation outputs for every arm in CLIMB. Every figure and table in the paper is computed from these files; none is stored pre-aggregated, so a reported number can be recomputed from the per-molecule predictions.

Layout

<wave>/<run>/moleculenet_cv/          5-fold scaffold CV on BACE, Tox21, QM7, HIV
chemeleon_suite/moleculeace/<run>/    30 activity-cliff targets, 3 evaluation seeds
chemeleon_suite/polaris/<run>/        Polaris tasks, predictions plus scores
cbs_benchmark/<run>/moleculenet_cv/   CBS rare-actives virtual screen, benchmark's own folds
wong_saureus/<run>/, fartdb/<run>/    the two additional external suites
analysis_rigor/                       bootstrap and multiple-testing outputs

Files

File	Contents
moleculenet_summary.csv	one row per head seed and fold, all metrics computed for that dataset
suite_summary.json	dataset means over folds and seeds
test_predictions.csv	per-molecule out-of-fold predictions; the source of every error bar
polaris_scores.csv	Polaris scores from the official evaluator against held-out labels
results.csv	suite-level long table: task, seed, subset, metric, value
verified.json	written only when every task and seed for that cell completed

Polaris test labels are held out by the benchmark, so predictions are scored off-box by scripts/chemeleon_suite_score_polaris.py. A Polaris results.csv therefore carries no scores by design, and polaris_scores.csv is the scored artifact.

Related

• Code: github.com/leifsieben/CLIMB
• Encoders: lsieben/climb-encoders
• Results: lsieben/climb-results
• Pretraining data: lsieben/climb-pretrain-data

Citation

@misc{climb2026,
  title  = {Does Pretraining Teach Chemical Language Models Chemistry?},
  author = {Sieben, Leif and Zimmermann, Yoel},
  year   = {2026},
  note   = {Preprint, arXiv},
  url    = {https://github.com/leifsieben/CLIMB}
}

License

CC-BY-4.0.