NanoLoop: 4,683 tiered nanoparticle candidate structures with energies and per-atom forces

A confidence-tiered set of 4,683 distinct relaxed nanoparticle structures with coordinates, energies and per-atom forces, produced by a closed-loop structure-inference study on bimetallic nanoparticles (Ni-Pt, Pd@Pt) and small metal clusters. Tier A (25 records) is ground truth: two nanoparticles with atomic coordinates from atomic electron tomography, plus 23 metal clusters whose global minimum is established by density-functional calculation. Tier B (186) is the candidates the inference loop selected. Tier C (4,472) is candidates that passed the stability gate but were not selected. Structures are supplied as extended XYZ with metadata on the comment line, together with a line-oriented JSON record file, a flat CSV index, and SHA-256 checksums for every file. Each record carries the interatomic potential that judged it, the run it came from and the source-code commit. IMPORTANT: for tiers B and C the energies and forces are machine-learned, not density-functional. The set must not be used to train an interatomic potential, because the labels were produced by the very potentials such training would improve. See README.md for the full caveats. Downloading the structures. The archive nanoloop-dataset-v1.tar.gz (548,848,359 bytes) is stored as 27 consecutive parts, nanoloop-dataset-v1.tar.gz.part-00 to part-26, because of an upload size limit. Download all parts and reassemble with cat nanoloop-dataset-v1.tar.gz.part-* > nanoloop-dataset-v1.tar.gz. The reassembled file has MD5 0059489be3f23bfb4ee8126e4d1d5bfa. README.md, index.csv, records.jsonl, summary.json and CHECKSUMS.sha256 are provided unsplit.