| --- |
| library_name: transformers |
| tags: [robotics, opendm, dm05, xpolicylab] |
| --- |
| # Checkpoints |
|
|
| ## `opendm/sim_exam-piper-l1-m1-piper-joint-0` |
| |
| DM0.5 (OpenDM) full SFT on the sim-exam dual Piper L1 split (900 episodes, 1,636,232 frames), |
| initialized from `DM05-MEM`. Experiment M1: 15,000 steps x global batch 256 |
| (8 x H20, micro-batch 8, gradient accumulation 4), lr 5e-5, 1,000 warmup steps, |
| cosine to 2.5e-5. Action space: `piper` / `joint`. |
| |
| Files are the final export (inference weights, `config.json`, `norm_stats.json`, processor and |
| tokenizer, plus `training_config.json` / `training_dataset.json` / `trainer_state.json` for provenance). |
| Load the directory with the XPolicyLab `OpenDM` adapter. |
|
|
