Data for: Amortized Inference of Multi-Modal Posteriors using Likelihood-Weighted Normalizing Flows

Data accompanying "Amortized Inference of Multi-Modal Posteriors using Likelihood-Weighted Normalizing Flows" (R. Baruah, 2026), [arXiv:2512.04954 [cs.LG]]. This deposit contains the inputs and outputs of the 2D and 3D multimodal benchmark experiments of sections 4 and 5, and of the 4D B⁰ → J/ψ K⁰ CP-violation analysis of section 6. The analysis code is archived separately at [DOI of the software record] and developed at [GitHub URL]. Contents Reference posteriors — ground-truth samples for each benchmark target, used as the reference in every distance and divergence metric. LW-NF posterior samples — samples drawn from the trained flows, one file per base cardinality k and target, covering tables 2, 3 and 4. Weighted prior draws (benchmark_training_data.zip) — the likelihood-weighted training sets. Each row is a parameter vector followed by its likelihood weight. Only needed to retrain; the tables can be recomputed without it. 4D B-meson sweep — trained flow weights, 200,000 posterior samples per run, and a per-run meta.json recording the random seed, architecture, best validation loss and metrics for all twenty runs. MCMC reference chain — the converged Markov Chain Monte Carlo posterior against which the section-6 results are benchmarked. MANIFEST.md lists every file with its array shape and SHA-256; SHA256SUMS.txt allows integrity checking with sha256sum -c. Reproducing the published tables Download the loose reference and LW-NF sample files (about 15 MB in total; the training archive is not required) and run, from the software repository: python benchmarks/benchmark_metrics.py --data-dir . --manifest manifest_tables.json This recomputes the average marginal Wasserstein-1 distance, the sliced Wasserstein distance and the KL divergence for tables 2, 3 and 4. Note on reference3DPathological.json This file contains a substantial fraction of repeated rows, consistent with having been generated by resampling with replacement. Estimators that assume a continuous density are affected; see MANIFEST.md.