MetaDNS: Enhancing Exploration in Discrete Neural Samplers via Well-Tempered Metadynamics — ICML 2026 Reproduction Dataset

Pre-computed sample pickles, trained model weights, classical-baseline outputs (MCMC, Swendsen-Wang, Wang-Landau, WT-MetaD), and 18 Jupyter reproduction notebooks for the paper 'MetaDNS: Enhancing Exploration in Discrete Neural Samplers via Well-Tempered Metadynamics' (ICML 2026, https://openreview.net/forum?id=OY7Qe2ZSx9). Covers the Ising model (L=4,8,16), the Potts q=3 model (L=4,8,16), and the Cu-Au binary alloy (2x2x4 and 4x4x4 supercells). The MetaDNS source code is maintained separately at https://github.com/xiaochendu/metadns. See README.md for the data layout, pickle schema, and instructions.