pi05_corkinbox100_fullrel -- Checkpoint 5000

Partial training run, step 5000 / 20000. Training was interrupted at step 7500 due to a disk-full event on the training host (97 GB root, each checkpoint ~23 GB, save_freq=2500 with no retention). This is the last clean checkpoint.

Config

• Base model: lerobot/pi05_base
• Dataset: FabianKerj/corkinbox100-fullrel (100 eps, 37036 frames, 30 fps, dk1_follower, 2 cameras)
• Relative actions on all 7 DOF (joint_1..6.pos + gripper.pos)
• Batch 64, dtype bf16, gradient_checkpointing on
• LR 2.5e-05 (warmup complete at step 500)
• Step 5000 metrics: loss ~0.008, grad_norm ~0.17

Contents

• Root: pretrained_model files (model.safetensors, config.json, train_config.json, pre/post-processors)
• training_state/: optimizer / scheduler / rng state (needed to resume training)