This is the official repository for the mini-vec2vec paper. The method proposes a solution to the stability and cost problems of the vec2vec method for the unsupervised alignment of sentence embeddings.

Install the requirements (optionally inside a virtual environment):

pip install -r requirements.txt

This part generates the embeddings we are going to align. It is provided here for completeness.

If you want to generate the embeddings yourself, you can run:

python generate_embeddings.py --output_dir output_dir/ --batch_size 32 --encoders e5,stella,granite,gtr

This will create one file per encoder in output_dir/.

This step is not part of the alignment algorithm, of course, and can be safely ignored -- it just creates the data for the algorithm to run on.

The output should be the same as vectors as in my huggingface repo dar-tau/nq-embeddings.

This step is more intensive than the alignment algorithm itself, taking approximately one hour on one T4 GPU.

Open the notebook in linear_vec2vec.ipynb and follow the instructions. You can change the encoders in:

embed_A, embed_B = sent_embeds['e5'].cpu(), sent_embeds['gtr'].cpu()

@misc{dar2025minivec2vec,
      title={mini-vec2vec: Scaling Universal Geometry Alignment with Linear Transformations}, 
      author={Guy Dar},
      year={2025},
      eprint={2510.02348},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2510.02348}, 
}