Add ONNX model metadata

README.md

@@ -0,0 +1,35 @@
+---
+license: apache-2.0
+library_name: relational-transformers
+pipeline_tag: tabular-classification
+tags:
+  - relational-transformers
+  - onnx
+  - relational-data
+---
+
+# RT-J ONNX
+
+This repository contains the classification checkpoint from
+[`RelativeDB/rt-j-fp16`](https://huggingface.co/RelativeDB/rt-j-fp16), exported
+to ONNX for framework-neutral target prediction over caller-provided relational
+cell embeddings.
+
+Load and cache it automatically with:
+
+```python
+from relational_transformers import RelationalTransformer
+
+model = RelationalTransformer(backend="onnx")
+predictions = model.predict(batch)
+```
+
+The graph accepts the canonical `RelationalBatch` tensor fields. Batch size and
+cell count are dynamic; the text and column-embedding width is fixed at 384.
+Callers remain responsible for producing the model-ready cell embeddings and
+relations described in the
+[`relational-transformers` input contract](https://relationaltransformers.com/docs/relational_transformer/usage/batches.html).
+
+`model.onnx` is exported from the full published checkpoint. The release
+process verifies PyTorch and ONNX Runtime output parity at multiple dynamic
+context lengths before publishing the file.

config.json

@@ -0,0 +1,12 @@
+{
+  "architectures": ["RTJModel"],
+  "library_name": "relational-transformers",
+  "source_model": "RelativeDB/rt-j-fp16",
+  "task_type": "classification",
+  "model_file": "model.onnx",
+  "d_text": 384,
+  "d_model": 512,
+  "num_blocks": 12,
+  "num_heads": 8,
+  "dynamic_axes": ["batch", "cells"]
+}