Northwestern and Fermilab Release NEXUS Dataset for NVIDIA Ising Benchmark

Northwestern University and Fermilab released a high-dimensional dataset from the Northwestern EXperimental Underground Site (NEXUS) to serve as a primary test case for the NVIDIA Ising benchmark and the vision-language tuning agent NVQCA. The dataset was recorded during a month-long campaign at 107 meters depth and contains the first reported observations of correlated charge jumps across multiple qubits in a low-background underground environment. The collaboration pairs real experimental traces with synthetic data to produce a standardized benchmark for AI-driven quantum hardware diagnostics.

The dataset targets models that interpret 2D experimental plots such as sinusoidal interference fringes and sudden shifts caused by charge jumps. The benchmark evaluates models on six diagnostic tasks:

• •technical description
• •outcome classification
• •physical interpretation
• •fit quality assessment
• •parameter extraction
• •stability judgment

These tasks measure both visual pattern recognition and domain-specific inference, making the benchmark a vision-language evaluation rather than a pure numeric regression. The NVQCA agent is hosted on Fermilab's centralized GPU cluster and registered with a lab-wide model router, enabling researchers to submit experimental plots through a single endpoint and receive automated diagnostics in near real time. Combining real NEXUS data with synthetic cases improves task coverage and addresses class imbalance for rare events like multi-qubit correlated charge jumps.

This release formalizes a benchmark and dataset at the intersection of quantum hardware and AI-driven lab automation. Vision-language models that can read plots and explain hardware states are emerging as practical tools for calibration, drift detection, and failure diagnosis. The NEXUS dataset is notable because it provides real-world, low-noise measurements from an underground facility, reducing cosmic background that confounds surface experiments and revealing correlated noise modes that matter for superconducting qubits and error mitigation. Standardizing evaluation around the NVIDIA Ising benchmark and NVQCA creates comparability across model architectures, training regimes, and synthetic-data strategies.

Automated plot interpretation shortens the calibration loop for quantum processors and reduces manual expert bottlenecks. The benchmark's multi-task structure forces models to combine visual feature extraction with physics-aware reasoning, a useful proxy for other experimental sciences. The availability of production-grade GPU hosting and a model router demonstrates a deployment path from research benchmark to lab-integrated tooling.

Expect the dataset and benchmark to expand as the collaboration integrates additional testbeds under CosmiQ and LOUD programs and as groups contribute further synthetic scenarios. Key follow-ups will include open comparative results from alternative VLMs, detailed benchmark metrics on failure modes for correlated charge events, and research on domain adaptation for models trained on mixed synthetic and real experimental data.