| # Attribution and checkpoint scope |
|
|
| Phoenix-Pretrain distributes the Phoenix-MWA Robot37 Joint96 world-action |
| pretraining checkpoint at optimizer step 256013. It contains the video and |
| action experts and proprioception encoder, using 48 WanVAE channels and |
| 48 CS-VAE channels. It is a weights-only initialization checkpoint, not a |
| full optimizer-state resume or a benchmark post-trained policy. |
|
|
| The video expert is derived from Wan2.2-TI2V-5B. The Wan Team's upstream model |
| is distributed under Apache License 2.0; a copy is included in |
| `LICENSES/Apache-2.0.txt`. |
| Upstream: https://huggingface.co/Wan-AI/Wan2.2-TI2V-5B . |
|
|
| Phoenix Encoder and CS-VAE weights and their notices are distributed separately |
| in OpenPhoenix/Phoenix-Encoders and AnyverseDynamics/Phoenix-Encoders. |
| WanVAE, text/tokenizer components and training datasets are obtained separately |
| under their respective terms. |
|
|
| The public checkpoint uses Phoenix/CS-VAE identifiers and omits private local |
| paths. Those metadata changes preserve the model tensor names and values. |
| Retained upstream terms do not assign an additional license to Phoenix-owned |
| weight contributions. The Phoenix-MWA code license is not a blanket license |
| for model weights, datasets or third-party assets. Attribution does not imply |
| upstream endorsement. |
|
|
