Modelpublic

usejul/jul-decision-e5-small

sourceHugging Faceapache-2.0updated 8h agoView on Hugging Face

0likes

Model Card

jul-decision-e5-small

`multilingual-e5-small` fine-tuned to answer jul decisions in zero-shot. Same size and speed as e5-small; ONNX int8 build (98 MB): `usejul/jul-decision-e5-small-onnx`.

bash

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

Inputs are prefixed with query: (declared in config_sentence_transformers.json).

Results

	e5-small	jul-decision-e5-small
zero-shot, jul dev sets (n = 800)	0.532	0.611
tuned hybrid head (AG News, Banking77, Emotion)	0.782	0.790
SIB-200, 6 languages not trained on (task seen in en/fr)	0.639	0.839

Part of the zero-shot gain comes from task types seen in training (financial sentiment, topics). With task data (a Context or a tuned head) the two models are close.

Training

jul's own score as the loss, on public English and French classification datasets. Embeddings frozen, 1,000 steps on one GPU.

License

Apache-2.0. Based on multilingual-e5-small (MIT).