VGGT-Omega 1B-256 OxE+MimicGen+RoboCasa365 — step 23000 snapshot
Single checkpoint snapshot from the 3DA_unified VGGT-Omega 1B-256 multi-source
pretraining run, taken at training step 23,000.
Provenance
- Run:
[VGGTOMEGA]_oxe_mimicgen_robocasa365_H8_vggtomega256_text_8n_mb2acc2_compile_2315944
- W&B run id:
bags/robot-gld/ohf3lpoa
- Backbone: VGGT-Omega 1B (encoder_input_size=256, DINOv3 patch=16)
- Data: OXE 23-dataset mix + MimicGen + RoboCasa365 (Cosmos24 subset)
- Training: 8 nodes × 4 GH200 = 32 GPUs, CSCS Clariden
Important note: mixed-H training history
- Steps 0–22,000: trained with
history_horizon=4 (canonical run config)
- Steps 22,001–23,000: trained with
history_horizon=8 via a wrong-defaults resume attempt (job 2331403)
If you intend to use this for further H=4 training, the weights at step 22000
are cleaner. See the matching W&B run summary for per-step metrics.
File
0023000.pt — full DeepSpeed-style checkpoint
(~12.8 GB, contains model_state, optimizer_state, scheduler_state,
step, action/proprio normalizer stats)
