nvidia/NVIDIA-Ising-Calibration-1

NVIDIA Ising-Calibration-1

NVIDIA Ising-Calibration-1 is a vision-language model fine-tuned for quantum computing calibration experiment analysis. Built on Qwen3.5-35B-A3B, it reads qubit calibration plots and provides technical descriptions, experimental conclusions, parameter extraction, and experiment status assessment.

Model Details

• Base model: Qwen3.5-35B-A3B (MoE, 35B total params, 3B active per token, 256 experts, 8 active)
• Architecture: Qwen3_5MoeForConditionalGeneration
• Training: Supervised fine-tuning (SFT) on ~72K quantum calibration entries
• Dataset: Quantum calibration experiment images and analysis, augmented by Qwen3.5-397B-A17B-FP8
• Precision: BF16
• Context length: 262,144 tokens
• Supported modalities: Text + Image

Benchmark Results (Zero-Shot)

Evaluated on quantum calibration experiments across 6 task types:

Task	Qwen3.5-35B-A3B (base)	Ising-Calibration-1
Q1: Technical Description	85.4%	85.8%
Q2: Experimental Conclusion	50.0%	81.7%
Q3: Experimental Significance	42.1%	68.3%
Q4: Fit Reliability	28.6%	100.0%
Q5: Parameter Extraction	63.9%	68.1%
Q6: Experiment Status	61.9%	81.7%
Average	55.3%	80.9%

Supported Calibration Experiments

Rabi, Ramsey, T1, spectroscopy, GMM, DRAG, chevron, pinchoff, coupler flux, and more.

Usage

Serving with vLLM

vllm serve nvidia/NVIDIA-Ising-Calibration-1 \
  --tensor-parallel-size 2 \
  --data-parallel-size 4 \
  --enable-auto-tool-choice \
  --tool-call-parser qwen3_coder \
  --reasoning-parser qwen3 \
  --enable-prefix-caching \
  --enforce-eager \
  --max-num-seqs 128

Python (OpenAI-compatible API)

from openai import OpenAI

client = OpenAI(api_key="EMPTY", base_url="http://localhost:8001/v1")
response = client.chat.completions.create(
    model="nvidia/NVIDIA-Ising-Calibration-1",
    messages=[{
        "role": "user",
        "content": [
            {"type": "image_url", "image_url": {"url": "data:image/png;base64,..."}},
            {"type": "text", "text": "Analyze this calibration experiment."}
        ]
    }],
    max_tokens=2048
)
print(response.choices[0].message.content)

Training Details

• Method: 2-phase sequential SFT (full fine-tuning, vision tower frozen)
• Framework: LLaMA Factory 0.9.5 + DeepSpeed ZeRO-3
• Hardware: 8 nodes x 8 NVIDIA H100 80GB GPUs
• Precision: BF16
• Batch size: 128 (effective)
• Learning rate: 5e-6 (cosine schedule, 3% warmup)
• Epochs: 1
• Training time: ~35 minutes (Phase 2)

License

NVIDIA Open Model License