FinBERT Quantum Optimization Project

Overview

In current quantum computing experiments, pervasive noise and the lack of robust error correction mechanisms pose significant challenges to reproducibility and result stability. Because today’s quantum processors operate in the Noisy Intermediate-Scale Quantum (NISQ) regime without full error correction, various error sources – including qubit decoherence, imperfect gate operations (limited gate fidelity), inter-qubit crosstalk, and measurement (readout) errors – can accumulate throughout a computation​. Qubit decoherence refers to the gradual loss of quantum coherence as qubits interact with their environment, which causes quantum information to decay into classical states​. This limits how long computations can run before the quantum state essentially randomizes. Similarly, finite gate fidelities mean that each quantum logic gate has a non-zero chance of introducing an error due to control pulse imperfections or calibration drift. Crosstalk errors occur when operations on one qubit inadvertently disturb the state of a nearby qubit due to unintended coupling or control signal leakage and have been identified as a significant source of error in current hardware​. Meanwhile, readout errors during measurement can cause the reported qubit state to be incorrect (e.g., reading a 0 as 1 or vice versa), further skewing the results​. Collectively, these noise processes degrade the fidelity of quantum operations and lead to high variability in outputs. Running the same algorithm multiple times can yield fluctuating outcomes because random error fluctuations at each run accumulate differently, undermining reliable reproducibility​. This noise-induced variability can also disrupt algorithmic convergence – for instance, in iterative algorithms, noise can alter the measured objective function and prevent consistent convergence to the same solution. Overall, current quantum algorithm results often require statistical averaging over many runs to distinguish signal from noise, and even then the results have significant uncertainty. While these limitations currently constrain quantum computing experiments, ongoing advances in quantum error correction and error mitigation techniques are expected to gradually improve fidelity and stability, bringing more consistent and reproducible outcomes over time.

Project Structure

Folder/File	Purpose
/data/	Contains the dataset (all-data.csv) used for training.
/models/	Saved models after fine-tuning and quantum optimization.
/notebooks/	Jupyter notebooks for model fine-tuning (finetune.ipynb) and quantum optimization (QBert.ipynb).
/results/	Output figures like loss curves, confusion matrices.
.gitignore	To ignore unnecessary files in version control.
requirements.txt	List of libraries needed to run the project.
README.md	Project documentation (this file).

---

Technologies Used

• Python
• Hugging Face Transformers
• Qiskit
• PyTorch
• Pandas, NumPy, Matplotlib, Seaborn
• NLTK (Natural Language Toolkit)

---

How to Run

1. Clone the repository:

   git clone https://github.com/RahatulAshakin/FinBERT_Quantum_Optimization.git
   cd FinBERT_Quantum_Optimization

2. Install dependencies:

   pip install -r requirements.txt

3. Open Notebooks:

   • Open notebooks/finetune.ipynb to fine-tune FinBERT.
   • Open notebooks/QBert.ipynb to apply Quantum Optimization (QAOA, VQE).

4. View Results:

   • Confusion Matrix
   • Loss Curves
   • Classification Report

---

Results

Model	Accuracy
Classical FinBERT	~78%
FinBERT + QAOA Optimized	~89%

Quantum optimization led to ~11% improvement in classification accuracy compared to classical training.

Download Pre-trained Model

We host the large finbert_qaoa_optimized model externally to keep this repo lightweight.

You can download the models from the below link:

"https://drive.google.com/drive/folders/1IWYWJ9Eu9UMh04-B2Dzeyf353pK42uBS?usp=sharing"

---

Future Work

• Experiment with Grover's Algorithm for optimization.
• Scale up on larger financial datasets.
• Explore running QAOA/VQE on real quantum hardware.

---

License

This project is licensed under the MIT License.
Feel free to use, modify, and distribute it with attribution.

---

Author

[Md Rahatul Ashakin]
For any queries or collaboration, feel free to connect!