README.md

metadata

library_name: transformers
license: mit
datasets:
  - klue
  - dkoterwa/kor-sts
language:
  - ko
pipeline_tag: sentence-similarity

KF-DeBERTa-base-cross-STS

pre-trained model: KF-DeBERTa-base-cross-NLI {https://huggingface.co/deliciouscat/kf-deberta-base-cross-nli}

trained data:

• klue/sts: 1epoch
• dkoterwa/kor-sts: 2epoch

label scaling: 0~5 -> -1->1

bi-encoder STS 학습을 위한 dataset augmentation을 상정하고 훈련하였습니다. {https://arxiv.org/abs/2010.08240}

cosine similarity로 학습할 수 있도록 scaling 된 output이 추론됩니다.