nvidia/Ising-Calibration-1.5-31B-BF16
NVIDIA-Ising-Calibration-1.5-31B-BF16 is a 31 billion parameter dense multimodal vision-language model developed by NVIDIA, built upon the Gemma 4 31B architecture. It specializes in analyzing quantum computing calibration experiment plots and generating structured technical text across six categories. This model is optimized for automated or assisted calibration workflows in quantum computing research and development.
Loading preview...
NVIDIA-Ising-Calibration-1.5-31B-BF16 Overview
NVIDIA-Ising-Calibration-1.5-31B-BF16 is a 31 billion parameter dense multimodal vision-language model developed by NVIDIA. It is built on the Gemma 4 31B architecture and is specifically designed for quantum computing calibration plot understanding.
Key Capabilities
- Multimodal Input: Processes both text prompts and image inputs (RGB .png, .jpeg, .jpg) of quantum calibration experiment plots.
- Structured Technical Text Generation: Generates detailed technical analysis across six categories:
- Technical description
- Experimental conclusion
- Experimental significance
- Fit quality assessment
- Parameter extraction
- Experiment success classification
- Performance: Achieves a mean zero-shot score of 74.5 on the QCalEval benchmark, outperforming Gemma-4-31B-IT (68.8). In MM-ICL settings, it matches Gemma-4-31B-IT with a mean score of 81.2.
- Deployment: Ready for commercial use and served through NVIDIA NIM with a vLLM backend, supporting BF16 serving precision.
Use Cases
This model is ideal for:
- Quantum computing researchers and calibration engineers: Automating or assisting the analysis of quantum experiment plot images.
- Developers: Integrating advanced plot interpretation into quantum computing workflows to generate structured technical reports.
Training and Evaluation
The model was trained on a synthetic corpus of 72.5K entries, focusing on calibration plot interpretation. It was validated and evaluated using the QCalEval benchmark, a synthetic vision-language benchmark specifically designed for quantum calibration plots, assessing various interpretation and classification tasks.