raftbioworks/boltz2_accuracy_recovery_20260919_064306
Boltz2 accuracy recovery — ongoing research campaign Work in progress. This snapshot does not establish accurate 100k-residue folding. The campaign investigates numerical fidelity, symmetry compression, native-supervised adaptation and independent AF2 controls. It uses public small PDB assemblies and the unchanged Boltz2 checkpoint, plus explicitly identified learned adapters. Native coordinates are training labels and evaluation references; they are not supplied as inference… See the full description on the dataset page: https://huggingface.co/datasets/raftbioworks/boltz2_accuracy_recovery_20260919_064306.
Boltz2 accuracy recovery — ongoing research campaign
Work in progress. This snapshot does not establish accurate 100k-residue folding.
The campaign investigates numerical fidelity, symmetry compression, native-supervised adaptation and independent AF2 controls. It uses public small PDB assemblies and the unchanged Boltz2 checkpoint, plus explicitly identified learned adapters. Native coordinates are training labels and evaluation references; they are not supplied as inference templates. Symmetry-compressed predictions assume the supplied oligomeric group and ordering, derived from the deposited assemblies.
The adaptation panel initially contains 18 training, 7 validation and 13 frozen test assemblies. An additional 58 training assemblies pass the same global sequence-identity filter against held-out cases. This split is not guaranteed to be held out from pretrained Boltz/AF2 training, or to exclude distant structural homologs. Cases, filters and the locked test protocol are supplied.
Four feature-distillation runs completed 100,000 optimizer steps, saving every 5,000. Their combined feature-validation criterion selects the unchanged model. A pair-only variant improves native validation lDDT partially but has not established parity with stock. Native supervised adapter and LoRA experiments are still running; their final status will be published explicitly.
The arithmetic-repaired uncompressed adapter matches stock Cα coordinates bitwise on all 25 development cases in both empty-MSA and homolog-MSA conditions. The actual packaged patch additionally reproduces all atoms in four local controls, including deep MSAs. Compression remains approximate.
Cα correction: older prediction archives selected Boltz’s representative atom, usually Cβ. Use full atomic coordinates with token_to_center_atom, verified by atom names. Corrected tables supersede earlier native RMSD/lDDT/drift values. Earlier immutable prediction archives are in https://huggingface.co/datasets/raftbioworks/boltz2accuracyarchitecture20260919032813 and have a prominent correction notice.
The unchanged Boltz checkpoint is already archived at https://huggingface.co/datasets/raftbioworks/anthropicbigmodeaudit20260918232311 (`checkpoints/boltz2conf.ckpt, SHA-256 090e82ac8c92f5e943fa1b39e7410a44027bea7243c0bbb3caa67a77fc1428e1`). This dataset archives the unchanged AF2 Multimer v3 model-1 parameters used for the alternative-model controls. See LICENSES.md.
W&B: https://wandb.ai/gaorory-ucla-team/boltz-accuracy-recovery. A separate final report, raw result archives, data/feature caches, all additional training checkpoints, metrics and reproduction instructions will be added when the remaining jobs and locked evaluation finish. Do not interpret a checkpoint upload as model promotion.