hyper³labs

Your data is hierarchical. Your models should be too.

Welcome to hyper³labs (pronounced Hypercube Labs) — an applied research and engineering lab building competitive embedding models and tools to inspect, curate, and improve embedding and retrieval systems.

We believe that Euclidean space should be a choice, not a default. We develop representations and visualization methods that respect the intrinsic geometry of data, particularly hierarchical, taxonomic, and non-Euclidean structures.

🔮 Interactive Workbenches (Spaces)

We develop HyperView, an open-source workbench for understanding embedding spaces and dataset/model failures. Explore our live interactive demos:

• 🔮 HyperView Main Demo — Our core open-source workbench for visualizing high-dimensional embedding spaces, mapping taxonomy, and diagnosing model/dataset failures.
• 🐆 Jaguar Spot Re-ID — Spot-pattern diagnostic visualization for wildlife conservation. Part of our collaboration on individual identification and diagnostic verification.
• 📦 ABO Catalog Explorer — Visualizing taxonomies and fine-grained product hierarchies using the Amazon Berkeley Objects dataset.
• 🔍 VisA Industrial Search — Anomaly inspection and structural part finding in industrial manufacturing datasets.
• 👗 DeepFashion Search — Exploring zero-shot retrieval, attribute mapping, and structural hierarchies in e-commerce fashion datasets.
• 🎯 Precision Region Search — Interactive referring expression grounding demo using RefCOCOg and the Hyper3-CLIP model.

🧠 Open Models

• 🤗 hyper3labs/hyper3-clip-v0.5 — Our open-weight 512-dimensional Lorentzian (hyperbolic) vision-language checkpoint trained with compositional entailment constraints for hierarchy-sensitive retrieval.

📊 Datasets

• 🤗 hyper3labs/jaguar-re-id — Spot patterns dataset used for individual jaguar identification in the wild.
• 🤗 hyper3labs/amazon-berkeley-objects — Mirror of the Amazon Berkeley Objects dataset used for fine-grained hierarchical product search evaluation.
• 🤗 hyper3labs/jaguar-hyperview-demo — Stratified subset for spot-pattern verification and debugging.

📄 Publications

• Are We Recognizing the Jaguar or Its Background? A Diagnostic Framework for Jaguar Re-Identification
  Antonio Rueda-Toicen, Robert Vava, Matin Mahmood. arXiv:2604.09690 (April 2026).
  arXiv:2604.09690 | Hugging Face Dataset

Building the geometry of intelligence.