docs: explain trained localization model versions

README.md

@@ -0,0 +1,21 @@
+---
+library_name: opensportslib
+tags:
+- action-spotting
+- sports
+---
+# OSL-loc-NetVLAD
+
+OpenSportsLib localization models. `main` is the version catalog; load an explicit experiment revision using `LocalizationModel.from_pretrained`.
+
+<!-- osl-model-versions:start -->
+## Trained model versions
+
+Select a version below to read its recipe, data, measured metrics and usage instructions. Each weight/config pair belongs to the linked experiment and seed.
+
+| Version / README | Dataset | Seed | Scope | Measured test scores | Artifacts |
+|---|---|---:|---|---|---|
+| [soccernet_netvlad_resnet_pca512](https://huggingface.co/OpenSportsLab/OSL-loc-NetVLAD/tree/soccernet_netvlad_resnet_pca512) | SoccerNet-v2 | 1 | released-code PCA ablation | loose average-mAP: 48.3548%; tight average-mAP: 7.3769% | [weights](https://huggingface.co/OpenSportsLab/OSL-loc-NetVLAD/resolve/2fc2f7db8e08ca05515987107144f80fae487725/model.pth) · [config](https://huggingface.co/OpenSportsLab/OSL-loc-NetVLAD/resolve/2fc2f7db8e08ca05515987107144f80fae487725/config.yaml) |
+
+Per-seed measurements do not establish exact paper reproduction. Read each version's protocol and aggregation notes.
+<!-- osl-model-versions:end -->

model-versions.json

@@ -0,0 +1,13 @@
+{
+  "soccernet_netvlad_resnet_pca512": {
+    "card_url": "https://huggingface.co/OpenSportsLab/OSL-loc-NetVLAD/tree/soccernet_netvlad_resnet_pca512",
+    "config_url": "https://huggingface.co/OpenSportsLab/OSL-loc-NetVLAD/resolve/2fc2f7db8e08ca05515987107144f80fae487725/config.yaml",
+    "dataset": "SoccerNet-v2",
+    "experiment": "soccernet_netvlad_resnet_pca512",
+    "model_url": "https://huggingface.co/OpenSportsLab/OSL-loc-NetVLAD/resolve/2fc2f7db8e08ca05515987107144f80fae487725/model.pth",
+    "revision": "soccernet_netvlad_resnet_pca512",
+    "scope": "released-code PCA ablation",
+    "scores": "loose average-mAP: 48.3548%; tight average-mAP: 7.3769%",
+    "seed": 1
+  }
+}