vwam-oa-rel28-rh-cog-split-ema-50k-g4-bg-1031-r2
Public inference checkpoints from a user-owned V-WAM OpenArm training run.
Latest saved checkpoint captured at 2026-11-02T10:55:10.753168+00:00: 50000. This repository keeps one captured checkpoint per experiment. Use a folder only after its HF_INFERENCE_COMPLETE.json marker is present.
Weights are EMA, mapped from net_ema.* to net.* and cast to the corresponding serving tensor dtype by the existing V-WAM export_eval_minimal.py. No raw training net, optimizer, scheduler or trainer state is included. Native PyTorch DCP format is preserved; this is not a Transformers AutoModel or safetensors export.
Use the matching V-WAM runtime with the folder config.yaml, normalization.json, and checkpoints/iter_XXXXXXXXX path. Upstream frozen VLM/tokenizer/VAE assets referenced by the original config remain runtime dependencies. Recorded training paths may need to be overridden for a different machine.
Each export checks all tensor names/shapes/dtypes and sampled tensor values. Hub file sizes and hashes are verified before the completion marker is published. No robot rollout or policy success claim is made by this backup.