QCV-Dataset

132 Quantum Circuits · 5 Core Modalities · 792 Experiment Results · Bilingual Annotations

The first multimodal quantum circuit dataset for training and evaluating AI systems on quantum circuit understanding, code generation, and verification.

Dataset Summary

QCV-Dataset contains 132 quantum circuits across 13 categories, each with 5 core modalities: circuit diagram image, Amazon Braket SDK code, Qiskit code, simulation results (state vectors), and bilingual expert annotations. Additionally, 792 experimental model invocations (3 models × 2 prompting modes × 132 circuits) provide a comprehensive benchmark for evaluating visual AI agents on quantum code generation.

Dataset Structure

Config: circuits (default)

Feature	Type	Description
id	string	Unique circuit identifier (e.g., C01_deutsch_jozsa_3)
circuit_image	Image	Qiskit-generated circuit diagram (PNG, 150 DPI, IQP style)
braket_code	string	Amazon Braket SDK executable Python code
qiskit_code	string	Qiskit equivalent implementation
description_en	string	English algorithm description
description_cn	string	Chinese algorithm description
category	string	Circuit category (13 categories)
difficulty	string	Difficulty level: basic, intermediate, advanced
qubits	int32	Number of qubits (1–10)
gate_count	int32	Number of gates (or null)
depth	int32	Circuit depth (1–27)
blockchain_relevance	string	Blockchain relevance tag (if applicable)
state_vector_dim	int32	Dimension of state vector (2^qubits)
nonzero_amplitudes	int32	Number of nonzero amplitudes
state_vector_real	sequence[float64]	Real components of simulated state vector
state_vector_imag	sequence[float64]	Imaginary components of simulated state vector
target_description	string	Target task description
best_pass_rate	string	Best pass rate across all models (e.g., "5/6")
all_pass	bool	Whether circuit passed all model-mode combinations
all_fail	bool	Whether circuit failed all model-mode combinations

Config: experiments

Feature	Type	Description
circuit_id	string	Reference to circuit
model	string	Model name (claude-opus-4.6, claude-sonnet-4.6, claude-haiku-4.5)
mode	string	Prompting mode (bv = base vision, tv = thinking vision / chain-of-thought)
syntax_ok	bool	Whether generated code compiles
exec_ok	bool	Whether code executes without runtime errors
fidelity	float64	Unitary matrix fidelity score
pass	bool	Whether verification passed (fidelity >= 0.99)
error	string	Error message (if failed)

Config: failures

Annotated failure cases from model evaluation with error type classification.

Config: equivalences

Circuit equivalence pairs for verification benchmarking.

Categories (13)

ID	Category	Count	Qubits
demo	Basic Gates	5	1–3
inter	Intermediate	10	2–4
adv	Advanced Algorithms	6	3–5
blockchain	Blockchain Protocols	11	2–8
A	Gate Type Coverage	15	1–3
B	Qubit Scaling	12	4–10
C	Classical Algorithms	15	2–4
D	Variational/Parameterized	10	2–4
E	Error Correction	8	3–9
F	Quantum ML	10	2–8
G	Blockchain Extended	8	3–6
H	Visual Variants	10	2–4
I	BTC/Blockchain Security	12	4–7

Dataset Creation

Data Collection

• Circuit diagrams generated with Qiskit QuantumCircuit.draw("mpl", style="iqp") at 150 DPI with tight bounding boxes
• Ground-truth code implemented in Amazon Braket SDK
• All circuits verified executable on Amazon Braket LocalSimulator

Annotations

• Bilingual descriptions (EN/CN) created by domain experts
• Categories assigned based on algorithm type and complexity
• Difficulty levels determined by circuit depth and gate complexity

Experiment Results

Model	BV Pass%	TV Pass%	Credits/Correct
Claude Opus 4.6	78%	75%	0.778
Claude Sonnet 4.6	77%	75%	0.142
Claude Haiku 4.5	43%	46%	0.072

Key Findings:

• 45 circuits passed all 6 model-mode combinations
• 18 circuits failed all 6 combinations
• Structural complexity (not qubit count) determines success
• Chain-of-thought provides no benefit for strong models (delta = -3 to -4%) but modest improvement for weakest (delta = +5%)

Usage

Load the dataset

from datasets import load_dataset

# Load main circuits dataset
circuits = load_dataset("QuantBlockchain/qcv-dataset", "circuits", split="train")

# Load experiment results
experiments = load_dataset("QuantBlockchain/qcv-dataset", "experiments", split="train")

# Access a sample
sample = circuits[0]
print(sample["id"])              # C01_deutsch_jozsa_3
print(sample["circuit_image"])    # PIL.Image object
print(sample["braket_code"])      # Python code string
print(sample["description_en"])   # English description
print(sample["description_cn"])   # Chinese description

Filter by category

algo_circuits = circuits.filter(lambda x: x["category"] == "classical_algorithms")
small_circuits = circuits.filter(lambda x: x["qubits"] <= 3)
passing_circuits = circuits.filter(lambda x: x["all_pass"] == True)

Analyze experiment results

from collections import Counter

model_pass = {}
for exp in experiments:
    model = exp["model"]
    if model not in model_pass:
        model_pass[model] = {"total": 0, "passed": 0}
    model_pass[model]["total"] += 1
    if exp["pass"]:
        model_pass[model]["passed"] += 1

for model, stats in model_pass.items():
    rate = stats["passed"] / stats["total"] * 100
    print(f"{model}: {rate:.1f}% ({stats['passed']}/{stats['total']})")

Data Governance & Croissant

This dataset follows Croissant metadata standards for machine-readable dataset descriptions. The dataset card uses structured YAML front matter for discoverability and includes:

• Data provenance: Synthetic generation via Qiskit + expert curation
• Annotation methodology: Expert-generated bilingual descriptions
• Verification protocol: Unitary matrix fidelity >= 0.99 on Braket LocalSimulator
• Known limitations: Framework-specific (Braket SDK), simulation-only, EN/CN bilingual only
• Bias considerations: 23.5% blockchain-relevant circuits may skew toward cryptographic applications

The dataset also includes a Croissant-RAI (croissant-rai.jsonld) extension documenting responsible AI considerations, data limitations, and recommended use cases.

Limitations and Biases

Limitation	Description
Framework lock-in	Code is Amazon Braket SDK specific
Simulation gap	No hardware execution data; LocalSimulator results may differ from real QPUs
Language coverage	Bilingual EN/CN only
Depth range	1-27; may not represent extremely deep circuits
Domain skew	23.5% blockchain-relevant circuits over-represents cryptographic applications

Citation

@misc{liu2026qcv,
  title={QCV: Cost-Aware Evaluation of Visual AI Agents for Quantum Code Generation},
  author={Liu, Dongping and Zhang, Aoyu and Zhang, Luyao},
  year={2026},
  url={https://github.com/QuantBlockchain/quantum-circuit-vision}
}

License

MIT — see LICENSE

Additional Documentation

• DATASHEET.md — Full dataset documentation following Gebru et al. (2021)
• CITATION.cff — Machine-readable citation metadata
• CIRCUIT_CATALOG.md — Full listing of all 132 circuits