DGLD: Domain-Gated Latent Diffusion checkpoints for energetic-materials discovery

Trained model checkpoints accompanying the manuscript DGLD: Domain-Gated Latent Diffusion for the Discovery of Novel Energetic Materials. Includes two conditional latent denoisers (v3_best.pt, v4b_best.pt), the LIMO VAE encoder/decoder (limo_best.pt), two multi-head classifier-guidance score models (score_model_v3e.pt 5-head, score_model_v3f.pt 6-head with hazard), the SELFIES alphabet (vocab.json), and run metadata (meta.json: latent dim 1024, four conditional properties, normalisation stats, denoiser config). All weights are fp32 PyTorch state-dicts. See the manuscript §3-§4 for architecture and §5 for evaluation. Sampling reproduction: m1_bundle/m1_sweep.py in the companion code release.