Examples

These examples begin with application-owned values and embeddings, then pass model-ready cell vectors or a typed RelationalBatch to the transformer. They do not hide parsing or encoding inside the model API.

Install the example dependencies first:

python -m pip install -e ".[dev,onnx]" sentence-transformers

Prediction and analysis

• predict_issue.py — encode issue fields and make one classification prediction

• batch_predictions.py — score variable-length issue contexts in one call

• typed_customer_churn.py — build typed customer, order, and support tensors

• ablate_support_history.py — define and compare an explicit ablation

• evaluate_churn.py — run classification and ablation evaluators

Training

• tune_issue_head.py — fit and save a multiclass head over a frozen backbone

• finetune_churn.py — fine-tune all RT-J weights with mini-batches

Deployment

• export_onnx.py — export dynamic batch/sequence axes and verify ONNX Runtime

• inspect_meta_model.py — inspect model dimensions without allocating weights

• triton_fp8_inference.py — run native FP8 weights with Triton on CUDA

Run any file from this directory or from the repository root, for example:

python examples/batch_predictions.py

The prediction examples download RelativeDB/rt-j-fp16 and sentence-transformers/all-MiniLM-L12-v2 on first use. The Triton example requires a CUDA device and the triton extra.