VGGT-Omega 1B 256 OXE/MimicGen/RoboCasa365 Step 51000

Files

• 0051000.pt - consolidated checkpoint for resume/eval, saved before eval at step 51,000.
• config.yaml - resolved training config saved by the run.

Run

• W&B run id: 8gxiadnc
• W&B run name: [VGGTOMEGA]_hf23000_H4_cvlab1_8gpu_mb28acc1_zero2_ckptfirst_evalmb1_eval1000_20260522_003836
• Source local checkpoint path: /NHNHOME/WORKSPACE/0226010404_A/CVLAB/CVLAB1/jisang/3DA_unified/logs/train/[VGGTOMEGA]_resume8gxiadnc_officialteacher_ckpt34000_H4_cvlab1_8gpu_mb28acc1_zero2_eval1000_20260522_124402/robot-20260522-124541/checkpoints/0051000.pt
• Saved time: 2026-05-23 04:35:30 KST
• Checkpoint size: 2.91 GiB

Key settings

• Backbone: vggt_omega
• Teacher checkpoint: pristine VGGT-Omega text checkpoint (vggt_omega_1b_256_text.pt)
• Student resume base: step 34,000 from the same W&B run lineage
• H choices: [4]
• deep_temporal_causal_mask=true
• Dataset mixture: OXE / MimicGen / RoboCasa365
• Training batch: 8 GPUs, micro batch 28, grad accum 1, ZeRO-2
• Eval/checkpoint interval: 1000 steps

Step 51000 console snapshot

• train total: 0.3854
• train action L1: 0.0742
• train depth loss: 0.0386
• train rel@1/5/10: 0.402 / 0.755 / 0.889
• eval unified H4 L1 norm: 0.1940
• eval unified rel@5/10: 0.504 / 0.678
• eval samples: 104