Quantum-VQE-Mistral-GGUF
Model Description
This model is a fine-tuned version of a merged large language model specifically optimized for Quantum VQE (Variational Quantum Eigensolver) simulations. It is designed to generate syntactically correct OpenQASM 3.0 circuits for molecular simulations, such as the H2 molecule, considering hardware constraints like topology and gate errors.
Merge Strategy
The model was created using the DARE-TIES merge method via mergekit. It combines the strengths of general instruction following, code generation, and mathematical reasoning.
- Base Model:
mistralai/Mistral-7B-v0.3
- Component Models:
codellama/CodeLlama-7b-hf (Weight: 0.4, Density: 0.5)WizardLM/WizardMath-7B-V1.1 (Weight: 0.6, Density: 0.5)
Training Hyperparameters
The fine-tuning process utilized the following parameters:
- Max Sequence Length: 4096
- Quantization: 4-bit (load_in_4bit: True)
- Epochs: 1
- Fine-tuning Technique: LoRA/QLoRA via Unsloth
Evaluation & Accuracy
The model was evaluated on its ability to generate an ansatz for the H2 (Hydrogen) molecule simulation.
- Calculated Energy: -1.0680931640625002 Ha
- Reference Energy: -1.137 Ha
- Status: Evaluated for chemical accuracy targets.
Usage
This model is intended for use with quantum simulation frameworks to automate the generation of hardware-aware ansatz circuits.
