MNISQ: A Large-Scale Quantum Circuit Dataset for Machine Learning in the NISQ Era

Dataset Overview This Zenodo record contains the complete MNISQ (Machine Learning on Noisy Intermediate-Scale Quantum circuits) dataset. MNISQ is a large-scale collection of quantum circuits generated from classical image datasets using the Automatic Quantum Circuit Encoding (AQCE) method. The dataset comprises approximately 4.95 million samples and is designed for research in: - quantum machine learning - classical learning on quantum representations - noise-aware quantum computation - benchmarking of quantum and classical models on circuit data The dataset spans three source datasets: MNIST (`mnist_784`) Fashion-MNIST Kuzushiji-MNIST Each dataset is encoded into quantum circuits and released across multiple partitions, encoding formats, and fidelity levels. --- Archive Naming Convention Each archive follows the naming pattern: `[data]_[type]_[fidelity].zip` Where: data - `train_orig` = original encoded training split - `base_train_orig` = same split as `train_orig`, but in base OpenQASM format - `train` = augmented training split - `test` = encoded test split - `base_test` = same split as `test`, but in base OpenQASM format type - `mnist_784` - `Fashion-MNIST` - `Kuzushiji-MNIST` fidelity - `f80` = fidelity threshold greater than or equal to `0.80` - `f90` = fidelity threshold greater than or equal to `0.90` - `f95` = fidelity threshold greater than or equal to `0.95` Higher fidelity tiers correspond to more accurate approximations of the target quantum state and typically require deeper circuits. --- Dataset Composition The MNISQ dataset is organized as a family of ZIP archives covering: 3 source datasets 5 data partitions / encodings 3 fidelity tiers Total scale: approximately **4.95 million samples**. Each archive is independent but follows an identical schema, allowing selective download and seamless merging. --- Internal ZIP Structure Each ZIP archive contains one root folder named exactly like the archive and four aligned subfolders: - `qasm/` - `state/` - `label/` - `fidelity/` Samples are indexed by shared integer IDs: `0, 1, 2, ..., N-1` For a given index `i`, the files: - `qasm/i` - `state/i` - `label/i` - `fidelity/i` represent a single complete data sample. This structure guarantees deterministic alignment across all modalities. --- Per-File Content Format `label/i` - Plain-text integer class label - Example: `5` `fidelity/i` - Plain-text floating-point fidelity value - Example: `0.848595` `state/i` - Plain-text quantum state amplitudes - Format: `real_part imaginary_part` - One amplitude per line `qasm/i` - OpenQASM 2.0 circuit file Typical header: `OPENQASM 2.0;` `include "qelib1.inc";` `qreg q[10];` Two variants exist: - Dense QASM (with `DenseMatrix` gates, Qulacs-compatible) - Base QASM (standard gate decomposition, Qiskit/PennyLane compatible) --- Dataset Construction The dataset is generated by encoding classical image data into quantum states using the AQCE algorithm. Each sample corresponds to: - a classical input (image) - a target quantum state - a quantum circuit approximating that state Each circuit: - operates on up to 10 qubits - contains up to 100 two-qubit gates depending on fidelity - is optimized to maximize fidelity with the target state Fidelity measures the overlap between the generated state and the target state. --- Data Splits `train_orig` Original encoded training dataset without augmentation `train` Augmented training dataset using transformations such as rotation, cropping, and shifting `test` Held-out test dataset for evaluation `base_train_orig` and `base_test` Same datasets as above but converted to base OpenQASM format --- Fidelity Tiers The dataset includes three fidelity levels: `f80` Lower fidelity threshold, shorter circuits `f90` Intermediate fidelity `f95` High fidelity, deeper circuits and more accurate state approximation These tiers allow controlled experiments on accuracy versus circuit complexity. --- Dataset Scope and Packaging All archives in this Zenodo record: - share identical schema - use aligned indexing - use plain-text formats - are independently downloadable This enables: - partial dataset usage - scalable workflows - reproducible experiments --- How to Read One Sample For a given archive and index `i`: - Read `label/i` to obtain the class label - Read `fidelity/i` to obtain the fidelity value - Read `state/i` to obtain the quantum state vector - Read `qasm/i` to obtain the circuit --- Reproducibility and Design Guarantees The dataset follows a deterministic storage design: - shared numeric indexing - aligned modality folders - plain-text formats - fixed schema across all archives This ensures: - exact reproducibility - easy parsing and loading - consistent integration into machine learning pipelines --- Intended Use Cases The MNISQ dataset supports: - quantum machine learning benchmarking (QSVM, variational models) - classical sequence modeling on QASM (Transformers, S4, LSTM) - noise robustness and error mitigation studies - circuit complexity versus learnability analysis - hybrid quantum-classical research --- Notes - All archives belong to the same dataset family - Schema is uniform across all files - Users can combine archives seamlessly - Dataset design supports both quantum simulation and classical machine learning workflows