augustocsc/gpt2_medium_prefix_682k

Seriguela - Symbolic Regression with Large Language Models

A research project exploring the application of GPT-2 and other LLMs to symbolic regression through fine-tuning and reinforcement learning.

🏆 Key Results (Model Scaling Study - Feb 2025)

Model	Parameters	Valid Rate (Quality)	Valid Rate (Nguyen)	Avg R²	Max R²
Base	124M	99.4%	62.5%	0.9190	0.9994
Medium	355M	99.2%	75.2%	0.9812	0.9999
Large	774M	100% 🏆	89.0% 🏆	0.9852 🏆	1.0000 🏆⭐

Breakthrough: First model to achieve 100% valid rate and R²=1.0 perfect fit on Nguyen-8

---

📂 Project Structure

Organized by research phases for systematic experimentation:

seriguela/
├── 1_data/                 # FASE 1: Preparação de Dados
├── 2_training/            # FASE 2: Treinamento e Fine-tuning
├── 3_evaluation/          # FASE 3: Avaliação
├── 4_analysis/            # FASE 4: Análise e Visualização
├── models/                # Modelos treinados
├── results/               # Resultados de experimentos
├── docs/                  # Documentação completa
├── src/                   # Código fonte (package)
└── scripts/               # Scripts auxiliares

Each directory contains a README.md with detailed documentation.

---

🚀 Quick Start

# 1. Clone and setup
git clone https://github.com/augustocsc/seriguela.git
cd seriguela
python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt

# 2. Train model
cd 2_training/supervised
python train_with_json.py --model_size gpt2-medium

# 3. Evaluate
cd ../../3_evaluation/benchmarks
python run_all_nguyen_benchmarks.py --model_path ../../models/gpt2/medium_700k_json

# 4. Analyze
cd ../../4_analysis/visualization
python create_visualizations.py

---

📊 Complete Documentation

• Scientific Report: docs/reports/SCIENTIFIC_REPORT_MODEL_SCALING.md
• Developer Guide: docs/guides/CLAUDE.md
• Model Cards: docs/model_cards/

---

🎓 Citation

@misc{seriguela2025,
  title={Scaling Laws for Symbolic Regression with LLMs},
  author={Augusto Cesar},
  year={2025},
  note={First 100% valid rate + R²=1.0 achieved}
}

---

Status: ✅ Production-ready | 📊 Publication-ready | Last Updated: 2026-02-04