🌌 QAC-L: Quantum Action Codec
Neuro-Symbolic Compiler Adapter for Qwen2.5-Coder-7B
A lightweight, deterministic bridge between natural language and quantum execution.
Developed by Pulsate Labs
🧠 What It Is & What It Does
In conversational quantum computing systems, asking a Large Language Model to generate raw Python/Qiskit code on the fly is highly unstable. It leads to API deprecations, syntax hallucinations, and execution safety risks.
This LoRA adapter solves that problem. By acting as a strict intermediary compiler, it restricts the LLM's output to a custom, highly-structured Quantum Domain-Specific Language (DSL). This output is easily parsed, validated, and executed by a classical software boundary (the "Bouncer" shield), ensuring 100% stable execution.
⚡ Supported Translations
1. Quantum Superposition / Randomness:
User Input: "Can you do a quantum coin flip using 3 qubits?"
DSL Output:[ACTION: RANDOM] [QUBITS: 3]
2. Quantum Chemistry (H₂ VQE):
User Input: "Hey, compute the ground state energy of hydrogen with a distance of 1.4 Angstroms."
DSL Output:[ACTION: VQE] [DISTANCE: 1.4]
⚙️ How It Works (System Architecture)
The architecture relies on a Neuro-Symbolic loop, separating the creative translation of natural language from the strict mathematical execution of the quantum circuit.
graph TD
A[User Natural Query] -->|String| B(Qwen-7B LoRA Adapter)
B -->|Translates to DSL| C{The Bouncer Parser}
C -->|Invalid or Unsafe| X[Reject & Re-Prompt]
C -->|Valid: ACTION: VQE| D[Local Qiskit AerSimulator]
D -->|Raw Measurements| E(AI Physics Interpreter)
E -->|Scientific English| F[User Output]
🔬 The Fine-Tuning Process
The model was fine-tuned on a synthetic dataset of 1,200 conversational-to-DSL pairs generated using randomized, physically realistic parameter variations ($N \in [1, 5]$ qubits, $D \in [0.5, 2.5]$ Å).
Using QLoRA (Quantized Low-Rank Adaptation) on dual T4 GPUs, the model's training loss dropped from 1.89 to 0.11, indicating stable convergence and flawless syntactic replication of the target DSL.
📊 Comparison with Similar Approaches
When building AI-Quantum interfaces, developers typically use one of two architectures. Here is why the QAC-L Neuro-Symbolic pipeline outperforms standard generation:
| Feature | Standard LLM Code Gen (Raw Qiskit) | QAC-L Adapter + Bouncer (Ours) |
|---|---|---|
| Method | LLM writes raw Python script containing Qiskit commands. | LLM compiles parameters into a strict, validated schema. |
| Syntax Stability | ❌ Poor. Generates deprecated functions or incorrect imports. | High. 100% stable. Classical Python handles the API calls. |
| Security / Safety | ❌ Low. Vulnerable to Prompt Injection (running arbitrary OS code). | High. Sanitized by regex parser. Malformed inputs are blocked. |
| VRAM Footprint | ❌ Heavy. Requires massive frontier models (GPT-4) for reliability. | Light. Runs efficiently on local consumer hardware (7B Q4). |
Built for the future of Embodied AI and Quantum Control