rokrokss/KURE-v1-onnx

KURE-v1 (ONNX, int8)

An ONNX export of nlpai-lab/KURE-v1 (revision 8b418a58), quantized for CPU inference with transformers.js and onnxruntime. It is the embedding model shizue-plot uses for its chat memory.

All credit for the model belongs to its authors, the NLP & AI Lab at Korea University; this repository only converts it. It is distributed under the same MIT license as the original.

How it was made

• Exported with optimum's feature-extraction task (last_hidden_state).
• Quantized with onnxruntime dynamic int8, per channel, keeping the FFN output projections of the last 12 layers in fp32 (721 MB; onnx/model_quantized.onnx, sha256 3530a7640919b643a1cdf07894dbb2481a4edaf5acd0e439b8b860026f87b494).
• Pooling is the original's: the CLS token, then L2 normalization (1_Pooling + Normalize in the sentence-transformers model). Mean pooling gives much worse vectors.

Cosine similarity to the original sentence-transformers model on the same sentences: Korean (22) min 0.9909, mean 0.9928; English (11) min 0.9922, mean 0.9941.

Usage (transformers.js)

import { pipeline } from '@huggingface/transformers';

const embed = await pipeline('feature-extraction', '<this repository>', { dtype: 'q8' });
const output = await embed(['눈이 내리는 밤', 'A snowy night'], { pooling: 'cls', normalize: true });

License and citation

MIT, as nlpai-lab/KURE-v1. If you use the model, please cite its authors:

@inproceedings{jang2025kure,
  title={KURE: Embedding Model for Korean-Specific Retrieval},
  author={Jang, Youngjoon and Son, Junyoung and Lee, Taemin and Hong, Seongtae and Park, JeongBae and Lim, Heuiseok},
  booktitle={Annual Conference on Human and Language Technology},
  pages={129--134},
  year={2025},
  organization={Human and Language Technology}
}