docs: HF model card — best-of-both front-matter derived from manifest

README.md

@@ -1,16 +1,24 @@
 ---
-tags:
-  - OpenRAL
-  - rskill
-  - gr00t
-  - nvidia
-  - vla
-  - franka
-  - libero
-  - manipulation
-license: other
 language:
-  - en
+- en
+license: other
+license_name: nvidia-open-model-license
+pipeline_tag: robotics
+tags:
+- OpenRAL
+- rskill
+- gr00t
+- vision-language-action
+- franka_panda
+- nvidia
+- vla
+- franka
+- libero
+- manipulation
+base_model:
+- nvidia/GR00T-N1.7-LIBERO
+base_model_relation: finetune
+inference: false
 ---
 
 # rskill-gr00t-n17-libero
@@ -19,22 +27,18 @@ language:
 > [LIBERO](https://libero-project.github.io/) benchmark, packaged for the
 > [OpenRAL](https://github.com/OpenRAL/openral) robot agent framework.
 
-This package bundles the **inference checkpoint** of
-[`nvidia/GR00T-N1.7-LIBERO`](https://huggingface.co/nvidia/GR00T-N1.7-LIBERO)
-at the repo **root** (config + `experiment_cfg` + statistics + safetensors;
-no DeepSpeed optimizer shards), plus a `rskill.yaml` manifest that adds
-capability checking, license surfacing, latency budgets, and local registry
-integration. The upstream repo keeps its weights in a `libero_spatial/`
-subfolder with no root `model.safetensors`, so this repackage is what makes
-the GR00T sidecar's `from_pretrained` load it directly. Weights are
-redistributed under the NVIDIA Open Model License (see `source_repo`).
-
-> **Runtime status.** Runtime-validated (ADR-0046). The out-of-process
-> adapter (`openral_sim.policies.gr00t`) + `tools/gr00t_sidecar.py` boot
-> helper auto-spawn a Python-3.10 GR00T sidecar that NF4-quantizes the
-> Cosmos backbone on load (~3 GB VRAM). Verified end-to-end on an RTX 4070
-> (8 GB): `episode_done` on `scenes/sim/libero_spatial`, ~100 ms/step, with
-> weights fetched from this repo via the manifest's `weights_uri`.
+This package wraps [`nvidia/GR00T-N1.7-LIBERO`](https://huggingface.co/nvidia/GR00T-N1.7-LIBERO)
+with a `rskill.yaml` manifest that adds capability checking, license
+surfacing, latency budgets, and local registry integration. It does **not**
+copy model weights.
+
+> **Runtime status.** This is a *packaging-and-validation* slice (ADR-0046
+> PR1): the manifest, license posture, and `model_family: gr00t` are wired
+> and tested. The out-of-process runtime adapter (`openral_sim.policies.gr00t`)
+> and the `tools/gr00t_sidecar.py` boot helper land in ADR-0046 PR2, which
+> also produces the locally-reproduced LIBERO eval numbers. Until then the
+> skill packages and validates but is gracefully dropped from a live policy
+> palette with an install hint.
 
 ## Upstream model, architecture & training
 
@@ -87,7 +91,7 @@ processor pipeline — hence no `processors` block in the manifest.
 | `embodiment_tags` | `franka_panda` |
 | `runtime` | `pytorch` (out-of-process sidecar, ADR-0046) |
 | `quantization.dtype` | `bf16` |
-| `weights_uri` | `hf://OpenRAL/rskill-gr00t-n17-libero` (this repo; root-level checkpoint) |
+| `weights_uri` | `hf://nvidia/GR00T-N1.7-LIBERO` |
 | `chunk_size` | 16 |
 | `state_contract.dim` / `action_contract.dim` | 8 / 7 |
 | `latency_budget.per_chunk_ms` | 1500 ms (sidecar round-trip + 3B inference) |