| library_name: pytorch | |
| tags: | |
| - motion-generation | |
| - dance | |
| # ACE Reverse Dance | |
| Intermediate snapshots from the Reactive and Responsive runs started on 2026-11-11. | |
| | Run | Published snapshot | Weights for inference | | |
| | --- | --- | --- | | |
| | Reactive | 7,605 optimizer updates, epoch 182 in progress | Online | | |
| | Responsive | Epoch 1,800 complete, 12,600 optimizer updates | EMA | | |
| Both files include online and EMA state dictionaries, the native normalizer, and snapshot configuration. | |
| Use the matching ACE runtime and codecs from the training run. The Reactive normalizer requires the runtime's | |
| `dataset` module when loading the native PyTorch checkpoint. Only load checkpoints from trusted sources. | |
| Both runs were still training at export time. The reported losses are training losses. The run receipts | |
| do not contain held-out metrics or an acceptance review of generated motion. See `snapshot-manifest.json` | |
| for provenance and SHA-256 checksums. | |