Questions-Gen: AI-Powered Mathematical Competition Problem Generation

Questions-Gen is a professional mathematical competition problem generation system based on Qwen3-14B, implementing a three-stage training strategy: Basic Pretraining → RL GRPO Optimization → Knowledge Distillation, specifically designed for generating high-quality mathematical competition problems.

🌟 Key Features

• 🎯 Three-Stage Training Pipeline: Basic Pretraining → RL GRPO Optimization → Knowledge Distillation
• 🔄 Intelligent Problem Variation Generation: Create smart variations of existing problems
• 📊 Multi-Dimensional Quality Assessment: Comprehensive problem quality evaluation system
• 🤖 Teacher Model Integration: Knowledge distillation from DeepSeek-R1
• 🚀 Ollama Integration: Convenient local inference deployment
• 📈 Batch Validation Tools: Large-scale model comparison testing
• 💎 Full Precision Models: Original FP16 precision without quantization loss
• ⌨️ Professional CLI Tools: Advanced command-line interface

🏆 Available Models

Training Stage	HuggingFace Model	Description	Downloads
Stage 1	xingqiang/QuestionsGen-Qwen3-14b-stage1-fp-merged	Basic mathematical problem generation	4+
Stage 2	xingqiang/questions-gen-qwen3-14b-stage2-merged-16bit	GRPO optimization + variation generation	3+
Final	xingqiang/questions-gen-qwen3-14b-final-merged-16bit	Complete knowledge distillation version	3+

🚀 Quick Start

Installation

# Install from PyPI
pip install questions-gen

# Install from source (development version)
git clone https://github.com/xingqiang/questions-gen.git
cd questions-gen
pip install -e .

Basic Usage

Quick Demonstrations

# Run complete functionality demo
python examples/quick_demo.py

# Model validation demo
python examples/demo_model_validation.py

# Ollama deployment demo
python examples/demo_ollama_push.py

Command Line Interface

# Validate final model
questions-gen validate --model final --tests 5

# Batch validation across categories
questions-gen batch --category algebra --tests 3 --export-csv

# Quality assessment
questions-gen quality "Find the derivative of f(x) = x³ + 2x² - 5x + 1" --detailed

# Ollama integration
questions-gen ollama --push-all
questions-gen ollama --test questions-gen-final

# HuggingFace tools
questions-gen hf --verify --compare

Model Training

# Custom training (requires GPU)
python scripts/questions_gen_training.py

Deployment Tools

# Quick import to Ollama
python tools/ollama_import.py

# Complete download and conversion
python tools/download_and_convert.py

Python API

from questions_gen import QuestionsGenTrainer
from questions_gen.validation import ModelValidator, BatchValidator, QualityEvaluator
from questions_gen.utils import OllamaManager, HuggingFaceUtils

# Model validation
validator = ModelValidator()
results = validator.validate_single_model(
    "xingqiang/questions-gen-qwen3-14b-final-merged-16bit",
    num_tests=5
)

# Batch validation
batch_validator = BatchValidator()
batch_results = batch_validator.comparative_batch_validation(
    category="calculus",
    tests_per_category=5
)

# Quality evaluation
evaluator = QualityEvaluator()
evaluation = evaluator.comprehensive_evaluation(
    "Prove that the square root of 2 is irrational."
)
print(f"Quality Score: {evaluation['overall_score']:.3f}")

# Ollama integration
ollama = OllamaManager()
ollama.push_all_models()

📊 Performance Results

Model Comparison

Based on comprehensive mathematical domain testing:

Model	Avg Quality Score	Generation Speed	Teacher Rating	Best Use Case
Final	0.847	2.1s	4.2/5.0	Professional competitions
Stage 2	0.782	1.8s	3.8/5.0	Problem variations
Stage 1	0.695	1.5s	3.4/5.0	Basic problem generation

Category Performance

Mathematical Domain	Final Model	Stage 2	Stage 1
Algebra	0.863	0.798	0.712
Calculus	0.891	0.815	0.698
Geometry	0.824	0.763	0.681
Number Theory	0.859	0.785	0.704

🔧 Advanced Features

Custom Model Training

from questions_gen import QuestionsGenTrainer, TrainingConfig

# Configure training parameters
config = TrainingConfig()
config.MAX_STEPS_STAGE1 = 100
config.PRESERVE_FULL_PRECISION = True

# Complete three-stage training
trainer = QuestionsGenTrainer()
trainer.train_full_pipeline()

Quality Assessment System

from questions_gen.validation import QualityEvaluator

evaluator = QualityEvaluator()

# Comprehensive evaluation
question = "Find all solutions to x⁴ - 5x² + 6 = 0"
evaluation = evaluator.comprehensive_evaluation(question)

print(f"Overall Score: {evaluation['overall_score']:.3f}")
print(f"Grade: {evaluation['grade']}")
print(f"Recommendations: {evaluation['recommendations']}")

Batch Processing

from questions_gen.validation import BatchValidator

batch_validator = BatchValidator()

# Multi-model comparison testing
results = batch_validator.comparative_batch_validation(
    models=[
        "xingqiang/questions-gen-qwen3-14b-stage2-merged-16bit",
        "xingqiang/questions-gen-qwen3-14b-final-merged-16bit"
    ],
    category="all",
    tests_per_category=3
)

# Export results
batch_validator.export_results_to_csv(results)

🐳 Ollama Local Deployment

Convenient local inference deployment:

# Install Ollama
curl -fsSL https://ollama.ai/install.sh | sh

# Push Questions-Gen models
questions-gen ollama --push-all

# Use the model
ollama run questions-gen-final "Generate a calculus competition problem:"

API Usage

import requests

def generate_problem(prompt):
    response = requests.post('http://localhost:11434/api/generate',
                           json={
                               'model': 'questions-gen-final',
                               'prompt': prompt,
                               'stream': False
                           })
    return response.json()['response']

problem = generate_problem("Create a number theory competition problem:")
print(problem)

🧪 Testing and Validation

Comprehensive Testing

# Model validation
questions-gen validate --model all --tests 5 --save

# Category-specific testing
questions-gen batch --category geometry --tests 5 --parallel

# Quality evaluation
questions-gen quality "Prove that √2 is irrational" --detailed

Model Comparison

# Compare all models
questions-gen compare --all-models --tests 5

# HuggingFace status check
questions-gen hf --compare --health

📈 Evaluation Dimensions

Quality Assessment Metrics

• Mathematical Content: Concept diversity and complexity
• Clarity: Problem statement clarity and structure
• Difficulty: Appropriate challenge level
• Completeness: Problem setup and constraints
• Originality: Innovation and creativity
• Educational Value: Learning objectives and pedagogy

Validation Categories

• Algebra: Equations, polynomials, abstract algebra
• Geometry: Euclidean, coordinate, solid geometry
• Calculus: Derivatives, integrals, optimization
• Number Theory: Primes, modular arithmetic, Diophantine equations
• Combinatorics: Counting, permutations, graph theory
• Analysis: Real analysis, sequences, convergence

🛠️ Development Guide

Project Structure

questions-gen/
├── questions_gen/          # Main package code
│   ├── cli/               # Command-line interface
│   ├── core/              # Core functionality
│   ├── data/              # Data processing
│   ├── models/            # Model components
│   ├── utils/             # Utility functions
│   └── validation/        # Validation system
├── docs/                  # Complete documentation
│   ├── guides/           # User guides
│   ├── technical/        # Technical documentation
│   └── training/         # Training-related docs
├── examples/             # Demo scripts
├── scripts/              # Training scripts
├── tools/                # Utility scripts
└── tests/                # Test files

Development Environment Setup

git clone https://github.com/xingqiang/questions-gen.git
cd questions-gen
pip install -e ".[dev]"

Running Tests

pytest tests/ -v --cov=questions_gen

Code Formatting

black questions_gen/
isort questions_gen/
flake8 questions_gen/

📚 System Architecture

Three-Stage Training Pipeline

Questions-Gen Training System
├── Basic Pretraining (Stage 1)
│   ├── Historical competition problems (50%)
│   ├── Conditional variations (30%)  
│   └── Innovative problem types (20%)
├── RL GRPO Optimization (Stage 2)
│   ├── Group policy generation (8 problems/group)
│   ├── Multi-dimensional reward function
│   └── Novelty constraint layer
└── Knowledge Distillation (Stage 3)
    ├── DeepSeek-R1 (difficulty prediction)
    ├── Logic rigor checking
    ├── Innovation assessment
    └── Educational value scoring

Reward Function System

reward = 0.4 * difficulty + 0.3 * novelty + 0.2 * rigor + 0.1 * diversity

• Difficulty Analysis (40%): Based on keywords and text complexity
• Innovation (30%): Difference from historical problems
• Logic Rigor (20%): Reasoning vocabulary density
• Diversity (10%): Within-group problem variance

📖 Documentation

• Complete Documentation: Documentation center entrance
• User Guide: Complete user manual
• Training Guide: Custom model training
• Technical Documentation: In-depth technical implementation
• API Reference: See docstrings in source code
• Example Code: Demo scripts and usage examples

🤝 Contributing

Contributions are welcome! Please see our Contributing Guidelines for details.

1. Fork the repository
2. Create a feature branch
3. Make your changes
4. Add tests
5. Submit a Pull Request

📄 License

This project is licensed under the Apache License 2.0 - see the LICENSE file for details.

🙏 Acknowledgments

• Unsloth: Efficient fine-tuning optimization
• HuggingFace: Model hosting and transformers library
• DeepSeek: Teacher model in knowledge distillation
• Qwen Team: Base Qwen3-14B model

📞 Support

• Issue Reports: GitHub Issues
• Model Downloads: HuggingFace Models
• Discussions: GitHub Discussions

🔗 Related Projects

• Unsloth - Fast LLM fine-tuning
• Transformers - State-of-the-art machine learning
• Ollama - Local LLM deployment

---

Questions-Gen - Advancing mathematical education through AI-powered problem generation. 🎓✨