LEAP Data Release

This repository contains data tables supporting the LEAP / Perovskite-RL manuscript. It is a data-release repository: model weights, Hugging Face training datasets, raw PDFs, logs, procurement tables, and the full internally curated candidate pool are not stored here.

Resource Links

Resource	Link	Notes
Manuscript	arXiv:2605.20242	Associated LEAP / Perovskite-RL preprint.
Model weights	JH976/Perovskite-RL	Perovskite-RL model repository. Update this link if the public model repository uses a different name.
Training datasets	datasets/JH976/Perovskite-RL	SFT and GRPO datasets for the language-model training stages.
Data-release repository	WD928/LEAP	Tables and source data used for benchmark, ablation, and candidate-selection reporting.

Repository Map

Module	Path	Contents
Hot-start additive data	data/	36 experimentally characterized additives, hard descriptors, soft mechanistic descriptor statistics, and relative PCE changes.
Mechanism benchmark	benchmark/questions.csv	32 multiple-choice questions from held-out literature sources.
Model benchmark results	benchmark/model_results/	Per-question answers for Perovskite-RL and baseline models.
Benchmark statistics	benchmark/statistics/	Accuracy summaries, exact McNemar-test tables, and Holm-Bonferroni-adjusted pairwise comparisons.
Representation ablation	ablation/representation/	Hard/soft/hybrid representation ablation source data and bootstrap confidence intervals for Figure 3 metrics.
Reasoning-source ablation	ablation/reasoning_source/	Perovskite-RL versus backbone soft-descriptor ablation tables and top-k diagnostics.
Decision-policy ablation	ablation/decision_policy/	Expected-improvement, predicted-mean, uncertainty, and random-policy comparison data.
Candidate selection	candidate_selection/	Round-specific top-50 validation shortlists with molecule identifiers and mechanism-score summaries.

Key Files

File	Description
data/hot_start_additives.csv	Main 36-additive hot-start table with measured PCE values and descriptors.
benchmark/statistics/model_summary_with_ci.csv	Benchmark accuracy summary with confidence intervals.
benchmark/statistics/mcnemar_vs_reference_holm.csv	Exact McNemar comparisons against Perovskite-RL with Holm-Bonferroni adjustment.
ablation/representation/figure3_bootstrap_ci_table.csv	Bootstrap confidence intervals for the Figure 3 representation-ablation metrics.
candidate_selection/top50_validation_shortlists_mechanism_scores.csv	Cleaned round-specific top-50 validation shortlist table.

License

This repository is released under the Apache License 2.0. See the LICENSE file for details.

Citation

Please cite the associated arXiv preprint if you use this repository:

https://arxiv.org/abs/2605.20242