pi05_telearms_v5_mixed

Intermediate checkpoint: step 21,750 of 39,063. Training is still running.

π₀.₅ (lerobot/pi05_base) full fine-tune on mliu48/telearms-yam-cube-v5-ablation-mixed (revision 20dca673): 60,093 simulated bimanual YAM demonstrations, a mixture of demonstrations of 7 tasks, 77.4M frames at 30 Hz, three cameras rendered like the real RealSense cameras (16:9) and letterboxed to 224×224.

One of three v5 ablations (high-score / high-Elo / mixed) trained with the same recipe and the same 40M-sample budget so they can be compared.

• Code: helen9975/RoboEval-telearms branch PI05_training @ 9dc9e9f (tools/training_diagnostics/train_v5.sh settings) plus local changes to training/train_pi.py (NaN guards, PyAV logging fix for data-loading speed), LeRobot 0.6.1 @ e40b58a8 with the repo's patch (tools/install_lerobot_patch.py).
• Full fine-tune (LLM + vision + action expert), bf16, gradient checkpointing; per-task prompt variants (meta/task_prompt_variants.json); 10% of each task held out for validation.
• 8 × H100 (1 node), batch 128/GPU (global 1,024); this checkpoint has seen 22.3M of 40M samples.
• AdamW (β = 0.9, 0.95, weight decay 0.01), peak LR 1.414e-4 (2.5e-5 × sqrt(1024/32)), 1,000 warmup steps, cosine decay to 1.414e-5 over 39,063 steps; 50-step action chunks; relative actions with absolute grippers.
• Training loss near this step: 0.030.
• These files are the checkpoint's pretrained_model/.
• Evaluation: not yet evaluated in simulation.