Pretrained Models

RelativeDB publishes RT-J checkpoints on the Hugging Face Hub. Each repository contains classification/ and regression/ subfolders with the same architecture and input contract, so one integration serves both task families.

from relational_transformers import RelationalTransformer

classifier = RelationalTransformer("RelativeDB/rt-j-fp16")
regressor = RelationalTransformer("RelativeDB/rt-j-fp16", task="regression")

Published Models

Model	Storage	Portable PyTorch	Triton CUDA	Intended use
`RelativeDB/rt-j-fp16`	16-bit	✓	✓	Highest-fidelity inference and fine-tuning
`RelativeDB/rt-j-fp8`	8-bit float	✓	✓	CUDA deployment with native FP8 weights
`RelativeDB/rt-j-int8`	8-bit integer	✓	—	Compact portable inference
`RelativeDB/rt-j-int4`	packed 4-bit	✓	—	Smallest portable checkpoint

Start with rt-j-fp16. It is the default checkpoint, the fine-tuning base, and the full-precision baseline to use when validating a deployment format.

Model Input Contract

The released models use 384-wide text and column embeddings from sentence-transformers/all-MiniLM-L12-v2, 512-wide hidden states, 12 blocks, 8 attention heads, and a 2048-wide feed-forward layer.

Matching only the embedding width is insufficient. A checkpoint is trained against one embedding space, one normalization scheme for scalars and timestamps, and the semantic type, target masking, and relation conventions described in Relational Batches. Callers must preserve all of them, and each model card records what its checkpoint expects. For a different encoder, see Changing the Embedding Space.

Classification and Ranking

The classification checkpoint emits a raw target logit. predict() applies a sigmoid by default. Request activation="identity" for ranking or calibration workflows that need the original logit; the sigmoid preserves ranking order, and calibration libraries expect the raw score.

Multiclass and multilabel tasks run through fitted task heads over the frozen backbone. The head tuning page shows the workflow.

Regression and Forecasting

Select task="regression" for scalar regression and forecasting. Predictions come back in the checkpoint’s normalized target space, so the application applies the inverse transform associated with its training statistics. Store that transform with the model artifact; a prediction without its denormalization is a bare number with no unit.