jay2219/Q-Route-Benchmark

⚛️ Q-Route Benchmark Dataset (qroute-benchmark-dataset)

The Q-Route Benchmark Dataset is a curated, post-training held-out evaluation suite comprising 100 non-overlapping quantum circuit routing test cases. It is specifically designed to audit and benchmark large language models (LLMs) on spatial hardware graph routing for physical Quantum Processing Units (QPUs).

---

📌 Dataset Summary

Physical quantum chips (such as IBM Heavy-Hex, linear chains, rings, or 2D grid lattices) restrict 2-qubit operations (cx, cz) to physically adjacent hardware qubits defined by a coupling map.

This dataset provides:

• Abstract Input OpenQASM 2.0 Circuits: Uncompiled circuits containing arbitrary multi-qubit interactions across non-adjacent physical qubits.
• Physical Hardware Specifications: Target QPU topology graphs, qubit counts, and explicit coupling map edge lists.
• Gold-Standard Ground Truth Outputs: Provably compliant, SWAP-routed OpenQASM 2.0 circuits with minimal SWAP overhead and circuit depth.

---

📊 Dataset Statistics & Topology Breakdown

The dataset contains exactly 100 benchmark test cases balanced across 5 distinct hardware graph topologies and 3 difficulty tiers:

Hardware Topology	QPU Architecture	Physical Qubits	Coupling Edges	Test Case Count	Difficulty Tier
Ring-8	Closed 1D Ring Lattice	8	16	20	Easy
Linear-9	1D Line Chain	9	16	20	Medium
Grid-4x3	2D Rectangular Grid Lattice	12	34	20	Medium
Star-15	Star / Central Hub Routing Graph	15	28	20	Hard
HeavyHex-16	IBM Heavy-Hex Connectivity Lattice	16	36	20	Hard

Total: 100 benchmark circuits | 5 hardware topologies | 3 difficulty tiers (Easy: 5–8 qubits, Medium: 9–12 qubits, Hard: 13–16 qubits).

---

🗂️ Data Structure & Schema

Each sample in quantum_benchmark.jsonl follows standard OpenAI Chat JSONL formatting with explicit metadata fields:

{
  "messages": [
    {
      "role": "system",
      "content": "You are Q-Route, an enterprise-grade deterministic Quantum Hardware Compiler..."
    },
    {
      "role": "user",
      "content": "### Physical Hardware Specification:\n- Target Topology: Grid-4x3\n- Number of Qubits: 12\n- Valid Physical Edges (Coupling Map): [[0, 1], [0, 3], ...]\n\n### Abstract Input OpenQASM 2.0:\n```qasm\nOPENQASM 2.0;\ninclude \"qelib1.inc\";\nqreg q[12];\n...\n```"
    },
    {
      "role": "assistant",
      "content": "OPENQASM 2.0;\ninclude \"qelib1.inc\";\nqreg q[12];\n...\nswap q[2], q[1];\nswap q[1], q[4];\ncx q[4], q[7];\n..."
    }
  ],
  "metadata": {
    "num_qubits": 12,
    "topology": "Grid-4x3",
    "num_edges": 34
  }
}

Key Fields:

• messages[0] (system): Verbatim prompt template system message enforcing strict OpenQASM compiler output.
• messages[1] (user): Complete prompt string containing target topology, qubit count, coupling map JSON, hardware constraints, and input abstract OpenQASM 2.0.
• messages[2] (assistant): Ground truth, topology-compliant OpenQASM 2.0 output with minimal SWAP gate insertions.
• metadata: Machine-readable metadata dict containing num_qubits, topology, and num_edges.

---

🔒 Test Set Integrity & Audit Trail

The 100 benchmark test cases in this repository were generated AFTER training data collection was finalized.

• Zero Overlap: Shared 0% overlap with the 15,000-sample training dataset (adaptions/qroute-training-data).
• Post-Training Generation: Guaranteed post-training test set generation to prevent data contamination or benchmark memorization.
• Verifiable Timestamps: Hugging Face commit history and GitHub repository creation timestamps provide an immutable audit trail.

---

🚀 How to Load and Use

Using datasets in Python:

from datasets import load_dataset

# Load the benchmark dataset from Hugging Face Hub
dataset = load_dataset("jay2219/Q-Route-Benchmark")

# Inspect a sample test case
sample = dataset["train"][0]
print("Topology:", sample["metadata"]["topology"])
print("Qubits:", sample["metadata"]["num_qubits"])
print("User Prompt:\n", sample["messages"][1]["content"])

📜 Citation & License

This dataset is released under the MIT License.

@dataset{qroute_benchmark_dataset_2026,
  author       = {Jay Prajapati},
  title        = {Q-Route Benchmark Dataset: 100 Held-Out Quantum Circuit Routing Evaluation Cases},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/jay2219/Q-Route-Benchmark}}
}