Hidden-State Privacy Has an Empty Middle (supplementary trained model checkpoints)

PyTorch checkpoint files accompanying the NeurIPS 2026 submission Hidden-State Privacy Has an Empty Middle. The archive contains 16 Split-Memory Transformer scaling-sweep checkpoints across 30M, 90M, 300M, and 1B parameters at fixed-token training with three seeds and two step variants at 90M, 8 SMT 90M-tier checkpoints for the four 12-layer architectures (baseline GPT, SMT main, SMT no-Jac, SMT $r=64$), and 88 predictive-quotient release checkpoints from the 44-cell variational-bottleneck negative-result sweep. JSON measurement and training logs for each checkpoint are released as a separate supplementary archive on OpenReview, with matching slugs. Code is at the anonymized repository linked in the paper. The archive layout mirrors the artifacts/{smt, smt_scaling, quotient_release} directories of the code repository.