Add README and model licence

README.md

@@ -0,0 +1,27 @@
+# ai-privacy-models
+
+Model artifacts used by the [Local Privacy Runtime](https://git.a75.it/andrearuffini/ai_privacy)
+(private repository). Only the files needed for local inference are published.
+
+## rizzo-pii-0.3B — revision d4f9893
+
+Published as the generic package
+[`rizzo-pii-0.3B` version `d4f9893`](https://git.a75.it/andrearuffini/-/packages/generic/rizzo-pii-0.3B/d4f9893).
+
+| File | SHA-256 |
+|---|---|
+| `config.json` | `6b9d5bdd4a194d40fb40c53fdc302e225514cd7940d97f657c4082454ee7e91b` |
+| `tokenizer.json` | `a0ab5772a8c5450f45a66c95985d6df2205bf09449ba7ff7bb46de034eaa118a` |
+| `model.onnx` | `32e403d1607f05acc2e1c2582418dbeec464c50ce7362eeacc9ea9dadce1dd3c` |
+
+- **Source model:** [`rizzoaiacademy/rizzo-pii-0.3B`](https://huggingface.co/rizzoaiacademy/rizzo-pii-0.3B),
+  revision `d4f989333bcda1fcb188785a12937b0f52b52580`, by Simone Rizzo — Rizzo AI Academy.
+- **`config.json`, `tokenizer.json`:** unchanged from the source revision.
+- **`model.onnx`:** FP32 ONNX export of the source weights, produced and validated by the
+  Local Privacy Runtime project (opset 18; parity with the PyTorch reference within
+  `atol=5e-5` on a synthetic corpus). Not an official artifact of the model author.
+- **Licence:** MIT, see [`LICENSE`](LICENSE).
+
+Do not trust these files because of where they come from: the runtime verifies every
+file against the SHA-256 values above, which are compiled into its binaries
+(`local-privacy-cli model install`).