vec2vec-workspace

goals

• confirm original repo vec2vec training is stable and reproducible
• extract original model loading setup
• extract eval setup for systematic eval against reproduction
• create minimal reproducible example for vec2vec training and eval

---

dev

1. librarify

• .checkpoints/ contains model checkpoints (including rjha18 pre-trained models @ final_model_release/)
• librarify vec2vec code for easier imports

2. extract eval setup setting up systematic eval comparison of translator

• unsupervised.toml line 17
• utils.py line 41 res_mlp
• train.py eval section line 523
• utils.eval_utils.py eval_loop line 336

notebook here

Original: Ever since the death of David Bowie back in 2016, notable celebrity deaths have dominated the Top 25 Reports
Generated (stella to gte then inverted): time, according to the sources quoted by the Ministry of Human Resources and Social Affairs. kwk was originally created as a response to
Generated (gte ground truth inverted): time, according to the sources quoted by the Ministry of Finance and the Ministry of Human Resources. Then-architect Josef t

3. reproduce

• minimal reproducible example for vec2vec training
• minimal eval example for vec2vec

---

vec2vec

notably on ubuntu 24.04 LTS i just had to cook up a new pyproject.toml so that i could use uv to manage the environment

in vec2vec/

uv sync --no-build-isolation
uv run python train.py unsupervised

# debugging
uv run python train.py unsupervised --num_points 10000 --epochs 10

# paper reproduction
uv run python train.py unsupervised --num_points 100000 --epochs 80

issues

• only trained on 1 GPU and is unstable for >1 gpu per issue
• num points https://github.com/rjha18/vec2vec/issues/14#issuecomment-2997194765

---

Citation

@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}, 
}