Quantum Physics Text Generator: Fine-tuning GPT-2 with PyTorch

A demonstration of transfer learning and generative AI applied to quantum physics domain. This project fine-tunes OpenAI's GPT-2 language model on quantum physics text to generate coherent, domain-specific passages. It bridges domain expertise in quantum information and computation with modern deep learning practices.

Author: Thiago Girao - PhD candidate in Physics, researching quantum information and quantum computing

---

Motivation

This project demonstrates how domain expertise in physics can be leveraged with state-of-the-art machine learning techniques. Rather than training a language model from scratch (computationally expensive and data-hungry), we use transfer learning to adapt a pre-trained model to specialized domain language. The quantum physics text generator serves as a practical example of:

• Transfer learning: Leveraging pre-trained models for downstream tasks
• Domain adaptation: Specializing general models to physics terminology and concepts
• Modern deep learning workflows: Proper training loops, learning rate scheduling, and evaluation metrics

---

Project Overview

Files

• quantum_text_generator.py: Main training and generation script

  • Custom PyTorch Dataset class for efficient data loading
  • Training loop with gradient accumulation and learning rate scheduling
  • Text generation with various sampling strategies
  • Original quantum physics training data

• evaluate_model.py: Model evaluation and comparison

  • Perplexity computation on held-out test data
  • Comparison between base GPT-2 and fine-tuned model
  • Multiple sample generation with qualitative analysis

• requirements.txt: Python dependencies

• .gitignore: Git ignore rules

• README.md: This file

---

Model Architecture

GPT-2: Transformer Decoder Language Model

GPT-2 is a transformer-based language model consisting of a decoder stack (left-to-right self-attention):

Input Text
    ↓
Token Embeddings + Positional Embeddings
    ↓
[Transformer Decoder Block × 12]  (for GPT-2 Small)
    ├─ Multi-head Self-Attention (12 heads)
    ├─ Feed-forward Network (4H hidden units)
    ├─ Layer Normalization
    └─ Residual Connections
    ↓
Output Layer Norm
    ↓
Linear Layer + Softmax
    ↓
Next Token Distribution

Key properties:

• Causal masking: Attention only to previous tokens (left-to-right language modeling)
• Parameterized: ~124M parameters for GPT-2 Small
• Pre-trained: On 40GB of diverse internet text (WebText)
• Transfer learning: Fine-tune on downstream tasks with modest computational resources

---

Fine-tuning Approach

Transfer Learning Strategy

We leverage domain-specific transfer learning with these practices:

1. Warm-start: Load pre-trained weights from OpenAI's GPT-2
2. Lower learning rate: 5e-5 (smaller steps than initial training)
3. Gradient accumulation: Simulate larger batch sizes without OOM errors
4. Gradient clipping: Stabilize training with max norm = 1.0
5. Learning rate scheduling: Linear warmup then decay
   • Warmup for 10% of training steps (avoids sudden weight changes)
   • Linear decay to small learning rate

Training Configuration

Optimizer: AdamW (weight decay = 0.01 for regularization)
Learning rate: 5e-5
Batch size: 2 (with gradient accumulation × 4)
Gradient accumulation steps: 4 (effective batch = 8)
Epochs: 3
Max sequence length: 256 tokens

Why These Choices?

• AdamW: Decoupled weight decay regularization prevents overfitting
• Low learning rate: Preserves pre-trained knowledge while adapting to quantum physics domain
• Gradient accumulation: Allows larger effective batch size on limited VRAM
• Gradient clipping: Prevents exploding gradients in transformer models
• Short training: 3 epochs sufficient for fine-tuning (full retraining would be many more)

---

Text Generation Strategies

The project demonstrates three important text generation methods:

1. Top-K Sampling

• Restrict sampling to K most likely next tokens
• Reduces low-probability nonsense
• Default: K=50

2. Nucleus Sampling (Top-P)

• Sample from smallest set of tokens with cumulative probability ≥ P
• More flexible than top-K (uses variable number of tokens)
• Default: P=0.95

3. Temperature Scaling

• Control randomness of distribution
• Temperature = 1.0: unchanged distribution
• Temperature > 1.0: more random/diverse
• Temperature < 1.0: more conservative/focused
• Default: 0.8 (slightly conservative to maintain coherence)

These techniques together prevent both:

• Mode collapse (only generating one response)
• Incoherent gibberish (allowing too much randomness)

---

Training Data

The training data consists of 20 original quantum physics paragraphs covering:

• Quantum entanglement and Bell inequalities
• Variational quantum algorithms (VQE, QAOA)
• Many-body localization and thermalization
• Quantum error correction and topological codes
• Quantum simulation and condensed matter systems
• Quantum machine learning applications
• Quantum chaos and out-of-time-order correlators
• Adiabatic quantum computing
• Quantum phase transitions
• Quantum metrology and sensing
• Quantum key distribution and cryptography

All text is original, written to demonstrate authentic quantum physics domain knowledge and to maintain focus on the ML techniques rather than copyright issues.

---

Evaluation Metrics

Perplexity

Definition: Perplexity = exp(average cross-entropy loss)

• Lower perplexity indicates better model predictions
• Measures how surprised the model is on unseen data
• Comparison: Fine-tuned model vs. base GPT-2 on held-out quantum physics text

Interpretation:

• Base GPT-2: Untrained on quantum physics → higher perplexity
• Fine-tuned: Adapted to quantum domain → lower perplexity
• Improvement: Quantifies benefit of domain-specific fine-tuning

Qualitative Evaluation

Generated samples are evaluated for:

1. Domain relevance: Does output use quantum physics terminology?
2. Coherence: Are sentences grammatically correct and fluent?
3. Factual plausibility: Could this appear in a physics paper?
4. Diversity: Do multiple samples show variety or repetition?

---

How to Run

Installation

# Create virtual environment
python -m venv venv
source venv/bin/activate  # On Windows: venv\Scripts\activate

# Install dependencies
pip install -r requirements.txt

Training

# Fine-tune GPT-2 on quantum physics text
python quantum_text_generator.py

This will:

1. Load pre-trained GPT-2 and tokenizer
2. Prepare quantum physics training data
3. Train for 3 epochs with proper learning rate scheduling
4. Save fine-tuned model to ./quantum-gpt2/
5. Generate example outputs

Expected runtime: ~5-15 minutes (depending on hardware; GPU recommended)

Evaluation

# Compare models and generate samples
python evaluate_model.py

This will:

1. Load base GPT-2 and fine-tuned model
2. Compute perplexity on held-out test data
3. Generate multiple samples from various prompts
4. Print detailed evaluation report

---

Key Technologies

Component	Technology	Purpose
Model	GPT-2 (Transformer)	Decoder-only language model
Framework	PyTorch	Deep learning framework
NLP Library	Hugging Face Transformers	Pre-trained models and training utilities
Optimization	AdamW + LR Scheduling	Training algorithm and scheduling
Tokenization	GPT-2 BPE Tokenizer	50K subword vocabulary
Hardware	CUDA (GPU) or CPU	Hardware acceleration (optional)

---

Architecture Highlights

Custom Dataset Class

class QuantumPhysicsDataset(Dataset):
    """Efficient tokenization and batching for language model fine-tuning."""

Handles:

• Tokenization with padding and truncation
• Efficient batch loading
• Proper label handling for causal language modeling

Training Loop Best Practices

- Gradient accumulation (memory efficiency)
- Gradient clipping (stability)
- Learning rate scheduling (convergence)
- Progress tracking with tqdm
- Proper device handling (GPU/CPU)

Flexible Generation

def generate_text(..., temperature, top_k, top_p, ...):
    """Configurable sampling strategies for diverse outputs."""

Supports multiple sampling strategies and fine-grained control over generation quality.

---

Results

Example outputs (fine-tuned model):

Prompt: "Quantum entanglement is"
Generated: "Quantum entanglement is a fundamental resource in quantum information processing
where the quantum state of a composite system cannot be described as a product of independent
states. This correlation structure enables protocols like quantum teleportation and distributed
quantum computing across separated nodes."

Prompt: "The variational quantum eigensolver"
Generated: "The variational quantum eigensolver combines quantum circuits with classical
optimization to find ground state energies efficiently. By variationally preparing ansatz states
and evaluating expectations on near-term devices, VQE enables chemistry simulations on
current quantum hardware."

Fine-tuned perplexity typically improves by 15-30% on quantum physics text compared to base GPT-2.

---

Future Improvements

1. Larger training dataset: Collect more quantum physics abstracts/papers
2. Domain-specific tokenizer: Train BPE tokenizer on physics vocabulary
3. Longer context: Increase max_length for multi-paragraph generation
4. Conditional generation: Generate abstracts given titles
5. Prompt engineering: Develop better prompts for specific physics tasks
6. Quantization: Compress model for deployment
7. Evaluation metrics: BLEU, ROUGE for comparison with reference texts
8. Comparison models: Fine-tune GPT-3.5, LLaMA for baseline comparison

---

References

Papers

• Attention Is All You Need: Vaswani et al., 2017

  • Introduced the Transformer architecture underlying GPT-2
  • arXiv:1706.03762

• Language Models are Unsupervised Multitask Learners: Radford et al., 2019 (GPT-2 Paper)

  • Demonstrated generative pre-training at scale
  • OpenAI Blog

• BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding: Devlin et al., 2019

  • Complementary approach (bidirectional) to causal language modeling
  • arXiv:1810.04805

Resources

• Hugging Face Documentation: huggingface.co/docs
• PyTorch Documentation: pytorch.org/docs
• Neural Networks & Deep Learning (Goodfellow et al., 2016): Deep learning fundamentals

---

Contact & Attribution

Author: Thiago Girao Email: [thiagorgs@id.uff.br] Research Focus: Quantum information theory, quantum algorithms, quantum machine learning PhD Program: Physics (Quantum Information & Quantum Computing)

---

License

This project is provided as-is for educational and portfolio purposes.