jul-decision-e5-small-onnx

ONNX build of usejul/jul-decision-e5-small for jul on CPU: 8-bit weights, 4-bit embedding, 98 MB per part — the vectors at the root, the cross model in cross/.

jul models add jul-decision-e5-small --repo usejul/jul-decision-e5-small-onnx --backend onnx

Zero-shot on jul's dev sets: 0.609 (e5-small 8-bit: 0.536); 4 ms per text for the vectors, 4–7 ms per yes/no for the cross model, on one CPU. One part fits an AWS Lambda package with jul; jul pack says which one a set of questions needs. Guide: AWS Lambda. Results, training and license: see the main model card.

v1.0: the first build, vectors only (revision="v1.0").