RT-J Regressor (rt-j-reg)

The regression checkpoint of RT-J — a Relational Transformer foundation model for in-context / few-shot entity regression over multi-table relational databases (no per-task gradient training).

• Task type: regression (metric: MAE / Z-score nMAE ↓)
• Selected: SWA checkpoint at step 18,000 (best on validation MAE)
• Params: ~85.6M · dtype: bfloat16 · 12 blocks, d_model 512, 8 heads, d_ff 2048
• Files: model.safetensors (weights), config.json (dims + text-embedding model + provenance)

from rt.checkpoints import load_rt_model
model, config = load_rt_model("star-project/rt-j-reg", device="cuda")

Full model card, training details, evaluation, license, and citation: see the RT-J repository. Use star-project/rt-j-clf for classification.