Add attribution README — credit: PyTorch / torchvision

README.md

@@ -0,0 +1,56 @@
+---
+license: bsd-3-clause
+tags:
+- mirror
+- xai
+- video
+- action-recognition
+- kinetics-400
+---
+
+# R3D-18 Kinetics-400 (torchvision) — xaitalk mirror
+
+This is a **bit-identical mirror** of the canonical artifact from PyTorch / torchvision.
+
+The mirror exists only as a resilience fallback for the
+[xaitalk](https://github.com/xaitalk/xaitalk) library — the upstream remains
+authoritative. All credit and licensing for the model belong to the original
+authors.
+
+## Attribution
+
+| Field | Value |
+|---|---|
+| **Original authors** | PyTorch / torchvision |
+| **Upstream (authoritative)** | https://download.pytorch.org/models/r3d_18-b3b3357e.pth |
+| **Source repo** | https://github.com/pytorch/vision |
+| **Paper** | https://arxiv.org/abs/1711.11248 (Tran et al. 2018, A Closer Look at Spatiotemporal Convolutions) |
+| **License** | `bsd-3-clause` (inherited from upstream — please respect upstream's terms) |
+| **Mirror file** | `r3d_18-b3b3357e.pth` |
+| **SHA-256** | `b3b3357ead25631ec9c57362ff2128a92d0427e01e2cd184951a44380c3f2e9d` |
+| **Size** | 133,546,016 bytes (127.4 MB) |
+
+## How xaitalk loads this file
+
+```python
+from xaitalk.hub import ensure_model
+weights_path = ensure_model("r3d18-kinetics400")
+# Tries the canonical upstream first; falls back to this xaitalk mirror
+# automatically if upstream is unreachable.
+```
+
+## Why mirror?
+
+xaitalk's research-grade reproducibility claim relies on every weight file
+being recoverable years from now. We mirror artifacts ≤ 2.5 GB under
+`xaitalk/*-mirror` so the pipeline survives upstream URL changes, repo
+renames, or deletions. Bit-level parity with the canonical is asserted in
+CI via `python -m xaitalk.hub verify-mirrors`.
+
+## Citation
+
+If you use this model, **please cite the original paper** (not the mirror):
+
+```
+https://arxiv.org/abs/1711.11248 (Tran et al. 2018, A Closer Look at Spatiotemporal Convolutions)
+```