🧠 mini_imo

A Lightweight IMO-Style Math Solver + AutoGrader Pipeline

Elegant • Minimal • Fully Reproducible • LLM-Powered

---

✨ Overview

mini_imo is a complete, minimal implementation of an IMO-style mathematics evaluation pipeline, featuring:

• 🔢 Math Solver (LLM-based or your tiny Transformer)
• 🧩 Short-Answer AutoGrader with equivalence checking
• 📚 Proof AutoGrader using IMO-style rubric {0,1,6,7}
• 🚀 End-to-End Evaluation Script
• 🧪 Ready-to-run sample benchmark
• 🔬 Optional: tiny GPT math model (PyTorch)

This project mimics modern math-evaluation pipelines used in LLM reasoning research.

---

🔥 Features

🎯 Short Answer Grader

• Extracts final answer
• Algebraic / numeric equivalence
• Strict grading (Correct/Incorrect)

📝 Proof Grader

• Four-level rubric:
  • Incorrect
  • Partial
  • Almost
  • Correct
• Scoring mapped to {0, 1, 6, 7}
• Judges correctness & completeness

🤖 Solver Options

• GPT-style LLM via OpenAI API
• Or your own mini_gpt_math.py model

📈 Benchmark Runner

• Reads JSONL file
• Solves → Grades → Produces CSV report
• Summary: accuracy & proof score