arc-v3podc-fullft-robotonly

Template-v3 flagship pretrain, Pod C: NEO-2B backbone, first-28 attach map, full finetune, robot-only (no VLM co-training), 28 of 40 backbone layers built (28..39 dropped, irreversible). 3-source corpus: OXE + AgiBot + RH20T. Global batch 4096 = 8 per GPU x 8 GPU x 8 nodes x accum 8, 75k steps, lr 2e-4 -> 2e-5, bf16.

What these files are

Each step<N>/ema_fp32.pt is the rank-0 fp32 EMA state dict (758 tensors, decay 0.99) -- the weights you evaluate or finetune from.

Optimizer state is NOT included. These are archive/eval artifacts, not resume points; a run resumes from its newest checkpoint, which is kept on the cluster and is not published here. Sidecars (engine_meta.pt, prompt_template.json, config.yaml) carry the provenance needed to rebuild the exact model.

The frozen backbone weights come from the NEO base model via neo_model_path; this repo carries the trained state, not the base.