README.md

---
language:
- en
license: mit
size_categories:
- n<1K
task_categories:
- other
tags:
- quantum-computing
- quantum-noise
- error-mitigation
- NISQ
- IBM-Quantum
- transfer-learning
- few-shot-learning
- physics
pretty_name: "Quantum Noise Transfer: Cross-Device Few-Shot Adaptation"
dataset_info:
  features:
    - name: circuit_id
      dtype: string
    - name: backend
      dtype: string
    - name: circuit_type
      dtype: string
    - name: n_qubits
      dtype: int32
    - name: T1_mean
      dtype: float64
    - name: T2_mean
      dtype: float64
    - name: readout_error_mean
      dtype: float64
    - name: cx_error_mean
      dtype: float64
    - name: noisy_distribution
      dtype: string
    - name: ideal_distribution
      dtype: string
    - name: x
      sequence: float64
    - name: y
      sequence: float64
  splits:
    - name: train
      num_examples: 170
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-00000-of-00001.parquet
---

# Quantum Noise Transfer: Cross-Device Few-Shot Adaptation Dataset

[![arXiv](https://img.shields.io/badge/arXiv-2604.24397-b31b1b.svg)](https://arxiv.org/abs/2604.24397)

> **Paper:** [Few-Shot Cross-Device Transfer for Quantum Noise Modeling on Real Hardware](https://arxiv.org/abs/2604.24397)
>
> **Authors:** Sahil Al Farib, Sheikh Redwanul Islam, Azizur Rahman Anik

## Dataset Description

A real-hardware quantum noise dataset collected from two IBM Quantum devices for studying cross-device transfer learning in quantum error mitigation. Each sample pairs a **noisy output distribution** (measured on real hardware) with the corresponding **ideal output distribution** (from noiseless simulation), augmented with device calibration features.

### Source Devices

| Device | Role | Samples |
|:---|:---|:---:|
| `ibm_fez` | Source (Backend A) | 85 |
| `ibm_marrakesh` | Target (Backend B) | 85 |
| **Total** | | **170** |

### Circuit Composition

| Circuit Type | Count | Purpose |
|:---|:---:|:---|
| Random | 40 | Structural diversity; generalization |
| Bell state | 15 | Two-qubit entanglement; CX error sensitivity |
| GHZ state | 15 | Multi-qubit entanglement; error accumulation |
| QFT | 15 | Layered gate accumulation; coherent errors |

All circuits use **2-5 qubits**, **depth 2-8**, and **8,192 shots** per execution.

### Device Calibration Comparison

| Property | ibm_fez (A) | ibm_marrakesh (B) | Delta |
|:---|:---:|:---:|:---:|
| T1 (us) | 142.4 | 192.8 | +35.4% |
| T2 (us) | 104.1 | 114.0 | +9.6% |
| Readout error | 0.0285 | 0.0335 | +17.5% |
| CX gate error | 0.0328 | 0.0560 | **+70.7%** |

## Dataset Fields

| Field | Type | Description |
|:---|:---|:---|
| `circuit_id` | string | Unique circuit identifier (UUID) |
| `backend` | string | IBM Quantum backend name (`ibm_fez` or `ibm_marrakesh`) |
| `circuit_type` | string | Circuit family: `random`, `bell`, `ghz`, or `qft` |
| `n_qubits` | int | Number of qubits (2-5) |
| `T1_mean` | float | Mean qubit relaxation time (seconds) |
| `T2_mean` | float | Mean qubit dephasing time (seconds) |
| `readout_error_mean` | float | Mean readout error rate |
| `cx_error_mean` | float | Mean CX (CNOT) gate error rate |
| `noisy_distribution` | string (JSON) | Measured probability distribution from real hardware |
| `ideal_distribution` | string (JSON) | Ground-truth distribution from noiseless simulation |
| `x` | list[float] | 41-dim input feature vector (circuit + calibration + noisy dist) |
| `y` | list[float] | 32-dim target vector (ideal distribution, zero-padded) |

### Feature Vector (`x`) Layout (41 dimensions)

| Index | Feature |
|:---|:---|
| 0 | Number of qubits |
| 1 | Circuit depth |
| 2 | CX gate count |
| 3 | H gate count |
| 4 | X gate count |
| 5 | Mean T1 (standardized) |
| 6 | Mean T2 (standardized) |
| 7 | Mean readout error (standardized) |
| 8 | Mean CX gate error (standardized) |
| 9-40 | Noisy output distribution (32-dim, zero-padded) |

## Additional Files

```
figures/              # All paper figures (PNG + PDF)
data/raw/             # Raw data from IBM Quantum hardware
  calibration_A.json  # ibm_fez calibration snapshot
  calibration_B.json  # ibm_marrakesh calibration snapshot
  circuit_meta.json   # Circuit structure metadata
  ideal.json          # Ideal (simulated) distributions
  noisy_A.json        # Noisy measurements from ibm_fez
  noisy_B.json        # Noisy measurements from ibm_marrakesh
data/processed/       # Preprocessed datasets
  dataset.json        # Full dataset (JSON format)
  dataset_standardized.json  # Standardized features
results/              # Experiment results
  experiment_results_fixed.json
  ablation_results_fixed.json
  training_history.json
  example_prediction.json
```

## Usage

```python
from datasets import load_dataset

ds = load_dataset("sahilfarib/quantum-noise-transfer")

# Filter by backend
source = ds["train"].filter(lambda x: x["backend"] == "ibm_fez")
target = ds["train"].filter(lambda x: x["backend"] == "ibm_marrakesh")

print(f"Source samples: {len(source)}")  # 85
print(f"Target samples: {len(target)}")  # 85
```

## Key Results

Using this dataset, a Residual Noise Adapter trained on `ibm_fez` achieves:

| Condition | KL Divergence | Improvement |
|:---|:---:|:---:|
| In-domain (A->A) | 0.3014 | -- |
| Zero-shot (A->B) | 1.6706 | baseline |
| Few-shot K=20 | **1.1924** | **-28.6%** |

## Citation

```bibtex
@article{farib2026fewshot,
  title={Few-Shot Cross-Device Transfer for Quantum Noise Modeling on Real Hardware},
  author={Farib, Sahil Al and Islam, Sheikh Redwanul and Anik, Azizur Rahman},
  journal={arXiv preprint arXiv:2604.24397},
  year={2026}
}
```

## License

MIT