AtomFormer: fixed inputs, model checkpoints, and molecular generation results

AtomFormer research data and model weights for ChEMBL37, MOSES and GDB13. Includes 27 AtomFormer checkpoints: supervised best, balanced reinforcement learning with anchor weight 0.10, and unanchored reinforcement learning, each with three training seeds per dataset. Fixed ChEMBL37/MOSES inputs and scored generation events cover 108 final evaluations: 54 validation-check and 54 test runs. Failed and unsupported attempts are retained. AtomFormer_Technical_Metadata_v2.0.0.zip contains per-run numerical metrics, official baseline model identities and pinned upstream source versions. MANIFEST.json records scientific archive checksums, input hashes and checkpoint provenance. LICENSES.txt preserves data licenses and upstream attribution. Official LibINVENT, PromptSMILES and DrugEx weights are referenced by source links and hashes. GDB13 parent structures must be obtained from the original provider under its terms. Selectors and annotations in https://doi.org/10.5281/zenodo.23468121 and the code reconstruction utilities recover the fixed inputs. GDB13 generated outputs and evaluation annotations are included. Code: https://github.com/Misaka10862/AtomFormer/tree/v2.0.0/v2 . Code license: GPL-3.0-or-later. Data licenses are separate. Scoring environment: RDKit 2022.09.5.