Vox

Applied AI: disaster-reconnaissance research, datasets, and developer tools.

Vox builds AI-powered products and tools. Here on Hugging Face we publish research checkpoints for unmanned-aircraft disaster reconnaissance, along with datasets for patent research and agent evaluation, including resources developed for Parallax. Explore our software and agent frameworks on GitHub.

Models published under this organization are developed in Japan by Vox, which operates from Tokyo.

Disaster-response models

Tancho studies how a drone could decide what to observe next after a flood or earthquake. Its current model proposes unanswered observation questions and semantic view requirements from imagery. Trusted deterministic software binds targets, checks constraints, and compiles any exact viewpoints and routes. A model response is not a flight command. In PX4/Gazebo simulation, the candidate-selection path now runs end to end, from the decision through the observation flight to the completion check, on procedural scenes. The published checkpoints are research artifacts; accuracy on real imagery, autonomous flight qualification, and Jetson Orin NX measurements remain open.

Training and evaluation sources differ by release. They include procedural synthetic scenes and limited public disaster UAV imagery; each model card states the source, measured results, corrections, and remaining limits.

• Tancho Observation Program v8 — A Cosmos3-Edge derivative for Reasoner and Generator roles. The model names unanswered questions and proposes semantic observation requirements. The repository root holds the v8 weights. Its 104/104 exact final result is from synthetic scenes, not an independent real-image benchmark.
• Tancho Cosmos Decision Preview — Experimental Jev-style candidate selection weights in the same repository. On one Modal L4, median decision-stage time on 12 real flood scenes, each tested in two candidate orders, was 0.5033 s versus 40.4724 s for v8 JSON generation (80.4x). This primarily measures the change from sequential generation to a single forward pass for candidate scoring, not a gain from training the new weights. Agreement with AI-review labels was 11/24 versus 12/24 for v8 using the same candidate mode; the preview has not shown an accuracy gain, and single-order choices in this candidate mode depend strongly on option position. The card also reports closed-loop simulation runs of this candidate path with the v8 parent weights (4 of 4 preregistered procedural slope missions completed).
• Tancho Generator v5 — A historical research baseline for numeric viewpoint generation, with different weights and output contracts from v8.

Tancho model weights are distributed under OpenMDW 1.1. Read the linked model cards for benchmark scope and license details.

Featured datasets

• Golden FTO · Layer A — Structured patent examination data linking inventions, cited prior art, and outcomes across the US, Europe, and Japan. Built to support Freedom-to-Operate (FTO) research and agent evaluation. The Japanese release contains derived labels only.
• OARD 2017 Mirror — A mirror of the USPTO Office Action Research Dataset used in the Layer A data pipeline for repeatable source-data access.

Explore and contribute

Read each model and dataset card for its scope, provenance, measurements, limitations, citation, and license terms. For questions, feedback, or data-quality reports, open a discussion in the relevant repository.

Website · GitHub · LinkedIn