Vec2Vec is a framework for training GANs (Generative Adversarial Networks) to convert between different embedding models. It allows for the transformation of embeddings from one latent space to another while preserving the semantic relationships between vectors.
Vec2Vec uses adversarial training to learn mappings between different embedding spaces. It can translate embeddings from unsupervised models like GTE (General Text Embeddings) to supervised models like GTR (General Text Representations), allowing for better alignment and utility of various embedding types.
Vec2Vec uses a TOML configuration file with sections for general settings, translator architecture, discriminator parameters, training hyperparameters, GAN-specific settings, evaluation metrics, and logging options.
To run the main unsupervised experiment:
python train.py unsupervisedThe train.py script with the unsupervised option accepts various parameters, including:
--seed: Random seed for reproducibility--sampling_seed: Seed for sampling operations--train_dataset_seed: Seed for training dataset generation--val_dataset_seed: Seed for validation dataset generation--normalize_embeddings: Whether to normalize embeddings--mixed_precision: Precision mode (e.g., 'fp16')--weight_init: Weight initialization method--dataset: Dataset to use (e.g., "nq")--max_seq_length: Maximum sequence length--unsup_emb: Unsupervised embedding model (e.g., 'gte')--sup_emb: Supervised embedding model (e.g., 'gtr')--n_embs_per_batch: Number of embeddings per batch--finetune_mode: Enable finetuning mode--noise_level: Level of noise to add
--translator
our paper is available on ArXiv: Harnessing the Universal Geometry of Embeddings (Jha, Zhang, Shmatikov, and Morris, 2025)
please cite it!
@misc{jha2025harnessinguniversalgeometryembeddings,
title={Harnessing the Universal Geometry of Embeddings},
author={Rishi Jha and Collin Zhang and Vitaly Shmatikov and John X. Morris},
year={2025},
eprint={2505.12540},
archivePrefix={arXiv},
primaryClass={cs.LG},
url={https://arxiv.org/abs/2505.12540},
}