raftbioworks/boltz2_quotient_validation_20260918_060803

Boltz 2 quotient-validation study Inference-only experiments with group-indexed pair features and tiled attention, compared with the supplied Anthropic optimization kit's actual big mode on H100 GPUs. Memory savings were measured; improved or preserved folding accuracy was not established. This is not a newly trained model. The checkpoint is the unchanged upstream Boltz 2 structure checkpoint, SHA-256 090e82ac8c92f5e943fa1b39e7410a44027bea7243c0bbb3caa67a77fc1428e1 (2,286,561… See the full description on the dataset page: https://huggingface.co/datasets/raftbioworks/boltz2_quotient_validation_20260918_060803.