Add model card explaining what these weights are and where they came from

README.md

@@ -0,0 +1,46 @@
+---
+license: other
+license_name: mixed-upstream
+license_link: https://github.com/deepsafehq/deepsafe-bench/blob/main/THIRD_PARTY_NOTICES.md
+tags:
+  - deepfake-detection
+  - model-zoo
+  - mirror
+---
+
+# DeepSafe Model Weights
+
+Weights for the 24 models used by
+[DeepSafe](https://github.com/deepsafehq/deepsafe-bench), a benchmark that
+tests how well deepfake detectors actually work.
+
+**These weights are not ours.** Each belongs to the authors of the model it
+came from and keeps that model's own license. See
+[THIRD_PARTY_NOTICES.md](https://github.com/deepsafehq/deepsafe-bench/blob/main/THIRD_PARTY_NOTICES.md).
+
+## Why mirror them
+
+Paper download links rot. DeepSafe lost one model outright when its weights
+were deleted from the host its paper pointed at. Mirroring means setup still
+works in three years.
+
+## Usage
+
+You do not normally download this by hand. `setup.sh` does it:
+
+```bash
+git clone https://github.com/deepsafehq/deepsafe-bench.git
+cd deepsafe-bench && bash setup.sh
+```
+
+Roughly 44 GB, covering 7 image, 9 video, 3 audio and 5 provenance models,
+plus the CLIP and face-detection caches they depend on. No token needed.
+
+Model *code* is mirrored separately at
+[deepsafe/model-code](https://huggingface.co/deepsafe/model-code).
+
+## Takedown
+
+If you authored any of these weights and want them removed, open an issue at
+https://github.com/deepsafehq/deepsafe-bench and they come down within 48
+hours, no questions asked.