Publish Phoenix-MWA evaluation resources (verified batch)

Explicit evaluation-only release; no training states, raw visual data or credentials.

README.md

@@ -0,0 +1,58 @@
+---
+library_name: pytorch
+tags:
+  - robotics
+  - phoenix-mwa
+  - dinov3
+  - visual-encoder
+---
+
+# Phoenix-Encoders
+
+Frozen visual-encoder checkpoints for **Phoenix-MWA**:
+
+| File | Checkpoint |
+| --- | --- |
+| `dino110k.pt` | DINO-derived encoder at step 110,000; Phoenix-MWA selects `ema_student` and backbone patch-token output. |
+| `stvae150k.pt` | Matching STVAE/CS-VAE encoder at step 150,000. |
+
+Use this pair with [Phoenix-LIBERO PT125k → FT25k](https://huggingface.co/OpenPhoenix/Phoenix-LIBERO) or [Phoenix-RoboCasa PT110k → FT45k](https://huggingface.co/OpenPhoenix/Phoenix-RoboCasa). These encoders are not complete robot policies and are not interchangeable with arbitrary DINO or VAE checkpoints. Phoenix-MWA provides the bundled encoder implementation and preprocessing configuration.
+
+## Download and use
+
+From the root of a local Phoenix-MWA checkout, install `huggingface_hub` and download the required repositories:
+
+```python
+from huggingface_hub import snapshot_download
+
+snapshot_download("OpenPhoenix/Phoenix-Encoders", local_dir="checkpoints/encoders")
+snapshot_download("OpenPhoenix/Phoenix-LIBERO", local_dir="checkpoints/libero")
+# For RoboCasa evaluation instead, download:
+# snapshot_download("OpenPhoenix/Phoenix-RoboCasa", local_dir="checkpoints/robocasa")
+```
+
+Follow the checkout's `docs/INSTALLATION.md` and `docs/RESOURCES.md`. Obtain matching Wan video/VAE, T5 text-encoder and tokenizer components separately under `checkpoints/base/`; they are not part of this encoder release. The policy configurations use [Wan2.2-TI2V-5B](https://huggingface.co/Wan-AI/Wan2.2-TI2V-5B) and the [Wan2.1-T2V-1.3B](https://huggingface.co/Wan-AI/Wan2.1-T2V-1.3B) tokenizer.
+
+For a LIBERO-Plus smoke run, first prepare the official benchmark resources at `resources/LIBERO-plus/libero/libero/`, containing `assets/`, `bddl_files/` and `init_files/`. Then:
+
+```bash
+python tools/prepare_eval.py --benchmark libero-plus \
+  --checkpoint checkpoints/libero/policy_phoenix_v1.pt \
+  --config checkpoints/libero/train_config.yaml \
+  --stats checkpoints/libero/dataset_stats.json \
+  --dino checkpoints/encoders/dino110k.pt \
+  --stvae checkpoints/encoders/stvae150k.pt \
+  --model-base checkpoints/base \
+  --benchmark-root resources/LIBERO-plus/libero/libero \
+  --output resources/phoenix-encoder-libero-smoke
+EVAL_GPUS=0 bash scripts/evaluate.sh libero-plus \
+  resources/phoenix-encoder-libero-smoke/eval.env.sh smoke
+```
+
+Use a new output directory for each run. The preparation step writes runtime configuration and does not launch inference. For RoboCasa, follow its linked model card, including `--metadata-root` for relocated package metadata and a separate simulator environment. Historical absolute paths in saved configurations/checkpoint metadata are provenance; use the explicit paths passed to `tools/prepare_eval.py` for the local run.
+
+Smoke checks execution rather than establishing a full benchmark score. PPU/NVIDIA numerical parity is not certified by this release.
+
+## Licensing
+
+See [NOTICE.md](NOTICE.md) and the accompanying license files, including the full **DINOv3 License**. DINOv3-derived materials retain those terms; Phoenix-MWA's source-code license does not replace them or assign an MIT license to all encoder weights. External Wan/text/tokenizer components and benchmark assets retain their own terms.

release_manifest.json

@@ -0,0 +1,59 @@
+{
+  "schema": "phoenix.huggingface_public_release.v1",
+  "repository": "OpenPhoenix/Phoenix-Encoders",
+  "code_project": "Phoenix-MWA",
+  "created_at": "2026-10-15T22:51:22.092274+08:00",
+  "files": [
+    {
+      "path": "LICENSES/DINOv3-LICENSE.md",
+      "size": 7503,
+      "sha256": "25d122eb8f5b880fd23c736fb6ea8018ee45c12237e00b8a86d14c653904999e",
+      "git_blob_sha1": "f531b1e6b5ab2318957bbf8ad1bda9f800a23e17"
+    },
+    {
+      "path": "NOTICE.md",
+      "size": 2012,
+      "sha256": "54cb4133a6054727b3f5ffe4436d035fa5b609e2747616e1238c9f0a20068910",
+      "git_blob_sha1": "63a272f06631a8562862727bc9490fa5ca4fc539"
+    },
+    {
+      "path": "README.md",
+      "size": 3355,
+      "sha256": "0f79ec603f09ac646c1d7442a50ab5c575dcfb77d788df2fa84101041602a032",
+      "git_blob_sha1": "0de154123afe2794b6c78d6303eb97d80e73d3a3"
+    },
+    {
+      "path": "dino110k.pt",
+      "size": 2485491825,
+      "sha256": "81a0082487879332968b5ec458a8d493bf30da601b855cfbaf0e09e250619053"
+    },
+    {
+      "path": "stvae150k.pt",
+      "size": 289373779,
+      "sha256": "7e41b75082444011bb6f6451c102e4a01bf2066faf20a7b0615fe05e0472d643"
+    }
+  ],
+  "excluded": [
+    "old_policy.pt",
+    "optimizer_state",
+    "scheduler_state",
+    "rng_state",
+    "pretraining_initialization_weights",
+    "raw_training_videos",
+    "visual_feature_caches",
+    "credentials",
+    "source_acceptance.json"
+  ],
+  "runtime_hash_checks_required": false,
+  "new_gpu_evaluation_performed_during_upload": false,
+  "encoder_provenance": {
+    "dino": {
+      "step": 110000,
+      "original_training_checkpoint_sha256": "ec83a00d2e65fab98a7f5c15a56700386dbf672dcde275ed7d99f16940650c0c"
+    },
+    "stvae": {
+      "step": 150000,
+      "original_training_checkpoint_sha256": "28d9b8efc16c2b7005e46dd9051d0c25429d64022d47492cd2e03b1470cf28f2"
+    }
+  }
+}