This repository contains the classification checkpoint from RelativeDB/rt-j-fp16, exported to ONNX for framework-neutral target prediction over caller-provided relational cell embeddings.

Load and cache it automatically with:

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.

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.