Measured small-protein ablations, an evaluation RNG/precision repair for the local Boltz quotient adapter, and a cyclic Fourier triangle-contraction prototype. No new weights were trained. This is not evidence of accurate 100k folding. See the report, 38 architecture ideas, and measurements. References are public PDB structures. checkpoint.json identifies the unchanged checkpoint already publicly archived in our preceding audit. results.tar.gz contains per-case/seed coordinates, input features, status, RNG hashes where recorded, and logs. Failed harness attempts are retained. The old sharedrng field denotes model-entry RNG only; use diffusionrngsha256 for the corrected comparison.…
Publicly accessible
mit
Joint protein representation autoencoder ablation. See config.yaml, provenance.json and status.json for the exact configuration and progress. W&B: https://wandb.ai/gaorory-ucla-team/frameflow-ltx/runs/e7pjq3aj Exact FP32 model weights are retained every 5 optimizer steps inside immutable checkpoints/.tar shards. Each member is a torch checkpoint containing vae, config, step, and trainseconds. checkpoints.jsonl records member names and SHA-256 hashes. Extract a member with Python tarfile, then load with torch.load. Full optimizer/RNG/data-position checkpoints are stored separately as resume.pt every 5000 steps and at termination; use these for exact resumption. This is autoencoder-only…
Publicly accessible
mit
