Skip to content

theaiagent/SynthEd

SynthEd: From synthetic data to simulated learners

GitHub release Python 3.10+ CI pytest Code style: ruff codecov DOI License: MIT

Agent-based simulation environment for Open & Distance Learning (ODL) research. SynthEd generates behaviorally grounded and temporally coherent learning trajectories by combining persona-driven agent modeling with 10 established theoretical frameworks. Built for researchers in learning analytics, educational data mining, and dropout prediction.

pip install -e ".[dev]"
python run_pipeline.py --n 200      # or: pip install synthedu

From statistical similarity to behavioral fidelity. Traditional synthetic data methods optimize for distributional match. SynthEd optimizes for behavioral coherence -- each data point emerges from a simulated student's evolving motivations, decisions, and life context.


Why SynthEd?

Challenge Traditional Approach SynthEd Approach
Privacy regulations (GDPR/KVKK) Anonymization (re-identification risk) Agents are fictional -- no real individuals
Class imbalance in dropout data Oversampling (SMOTE) -- loses context Parameter-level control of dropout rates
Temporal incoherence GAN/VAE post-hoc smoothing Persona + memory produces coherent trajectories

Key Features

Simulation Engine

  • 10 Theory Modules -- Tinto, Bean & Metzner, Kember, SDT, Garrison CoI, Moore, Rovai, Baulke, Epstein & Axtell, Gonzalez (+ unavoidable withdrawal mechanism)
  • TheoryModule Protocol -- 4-phase dispatch (individual, network, post-peer, engagement) with auto-discovery and _ENGAGEMENT_ORDER composition. New theories added with zero engine changes
  • Continuous Persona Spectrum -- Employment intensity, family responsibility, internet reliability as [0,1] floats with Beta distributions. No binary gates -- all theory effects scale continuously
  • Multi-Semester Simulation -- Carry-over mechanics for engagement, GPA, coping, dropout phases
  • GPA Feedback Loop -- Cumulative GPA anchors cost-benefit, non-fit perception, and competence beliefs

Calibration & Validation

  • Sobol Sensitivity -- 68-parameter sensitivity analysis identifying dominant dropout/engagement drivers
  • NSGA-II Calibration -- Multi-objective optimization with Pareto front, parallel --workers N support, adaptive parameter bounds
  • 5-Level Validation Suite -- 22 statistical tests (default; up to 24 with backstory validation) across distributions, correlations, temporal coherence, privacy, and backstory

Configuration

  • InstitutionalConfig -- 5 institution-level quality parameters that modulate theory constants. support_services_quality scales 13 Baulke dropout phase thresholds
  • GradingConfig -- Beta/Normal/Uniform grade distributions, dual-hurdle pass requirements, exam-only and continuous assessment modes, relative grading with t-score cohort normalization
  • EngineConfig -- 70 frozen engine constants with validation, overridable via dataclasses.replace()
  • PipelineConfig -- Frozen dataclass grouping 16 pipeline params with JSON serialization for reproducibility

Data & Integration

  • OULAD-Compatible Export -- 7-table CSV matching the Open University Learning Analytics Dataset schema
  • Optional LLM Enrichment -- Persona-grounded narrative backstories via OpenAI, Ollama, or any compatible provider
  • Benchmark Reports -- Customizable default profile with CLI report generation (--benchmark)

Quick Start

git clone https://github.com/theaiagent/SynthEd.git
cd SynthEd
pip install -e ".[dev]"              # Dev install (no LLM)
pip install -e ".[dev,llm]"          # Dev install with LLM support
python run_pipeline.py              # 200 students, 14 weeks
python run_pipeline.py --n 500      # Custom population
python run_pipeline.py --oulad      # OULAD-compatible export
python run_pipeline.py --benchmark  # Run default benchmark profile
python run_calibration.py --workers 4  # Parallel NSGA-II calibration
from synthed.pipeline import SynthEdPipeline
from synthed.pipeline_config import PipelineConfig

config = PipelineConfig(output_dir="./output", seed=42)
pipeline = SynthEdPipeline(config=config)
report = pipeline.run(n_students=300)
print(f"Dropout: {report['simulation_summary']['dropout_rate']:.1%}")

Use Cases

  1. Dropout Prediction -- Generate labeled training data with known ground-truth trajectories
  2. Intervention Simulation -- Model "what-if" scenarios by adjusting population parameters
  3. Privacy-Safe Benchmarking -- Share synthetic datasets publicly for reproducible research

Documentation

Document Content
User Guide Installation, configuration, calibration pipeline, OULAD export, LLM enrichment, troubleshooting
Theory & Architecture 10 theoretical anchors, factor clusters, architecture diagram, project structure, validation suite, test inventory

Roadmap

  • Multi-semester simulation with carry-over
  • 10 theory modules (Tinto, Bean & Metzner, Kember, SDT, Garrison, Moore, Rovai, Baulke, Epstein & Axtell, Gonzalez)
  • Trait-based calibration (Sobol + Optuna + OULAD validation)
  • Benchmark reports with CLI (--benchmark)
  • OULAD-compatible 7-table export
  • LLM enrichment with cost control and streaming
  • Disability severity (Beta distribution)
  • InstitutionalConfig (5 quality parameters modulating theory constants)
  • NSGA-II multi-objective calibration with Pareto front
  • GradingConfig (configurable grading policy: Beta/Normal/Uniform, dual-hurdle, exam-only)
  • EngineConfig (70 frozen engine constants with validation)
  • Relative grading (t-score cohort normalization)
  • PipelineConfig (frozen pipeline configuration with JSON serialization)
  • TheoryModule Protocol (phase-based dispatch with auto-discovery)
  • Engine modularization (state.py, grading.py, statistics.py -- engine.py 834→590 lines)
  • Engagement protocol unification (4th phase: contribute_engagement_delta)
  • Spectrum refactoring (binary → continuous for employment/family/internet)
  • GraphRAG integration (curriculum modeling)
  • LLM-augmented mode (forum posts, assignment text)
  • Parquet/Arrow export
  • PyPI package publication (pip install synthedu)
  • Interactive dashboard (Shiny + Plotly, dark/light themes, presets, validation)

Legal Disclaimer

SynthEd generates entirely fictional synthetic data. No real individuals are represented or identifiable. Outputs are intended for research, development, and educational purposes. SynthEd is under active development -- APIs and output formats may change between versions.

See full Legal Disclaimer and Responsible Use guidelines.


Contributing

Contributions welcome! See the User Guide for development setup.

ruff check synthed/ tests/
python -m pytest tests/ -v --tb=short

License

MIT License. See LICENSE.

Citation

If you use SynthEd in your research, please cite using the CITATION.cff file or the Zenodo DOI above.

Contributors

Contributor Role
Halis Aykut Cosgun Lead Developer, Data Scientist & AI Engineer, Researcher -- Yozgat Bozok University
Evrim Genc Kumtepe Research Advisor -- Anadolu University
Claude (Anthropic) AI pair programmer -- implementation, testing, code review

Acknowledgments

Conceptually inspired by TinyTroupe (Microsoft), MiroFish, and Agent Lightning. OULAD reference data: Kuzilek et al. (2017).