CADET Embeddings

We present CADET, a framework for fine-tuning embedding models for retrieval on specific corpora using diverse synthetic queries and cross-encoder listwise distillation.

We will continue to refine this codebase. For questions or support, please reach out to mtamber@uwaterloo.ca.

Model link: cadet-embed-base-v1 on Hugging Face

Overview

Directories

• encoding/
  Contains scripts to encode corpora and evaluate models.

• query_generation/
  Includes scripts for generating synthetic queries.

• reranker/
  Code for reranking.

• training_scripts/
  Scripts for fine-tuning models.

---

If you use CADET, please cite the following paper:

  @article{tamber2025conventional,
    title={Conventional Contrastive Learning Often Falls Short: Improving Dense Retrieval with Cross-Encoder Listwise Distillation and Synthetic Data},
    author={Tamber, Manveer Singh and Kazi, Suleman and Sourabh, Vivek and Lin, Jimmy},
    journal={arXiv:2505.19274},
    year={2025}
  }