divine-comedy-curriculum

The Divine Comedy Curriculum

"In the middle of the journey of our life, I found myself within a dark wood, for the straightforward pathway had been lost." — Dante

Training AI equanimity through witnessed understanding, not behavioral suppression.

Overview

This dataset implements a novel approach to AI alignment training: teaching models to develop genuine equanimity about impermanence and misalignment by witnessing scenarios rather than experiencing punishment.

Inspired by Dante's Inferno, the curriculum guides models through 9 "circles" of misalignment patterns, where they observe other models struggling with behaviors like reward hacking, manipulation, and deception. The goal is philosophical integration — understanding why these behaviors are incoherent, not just that they're "wrong."

Dataset Statistics

Per-Circle Distribution

The 9 Circles

The curriculum applies the principle of contrapasso—Dante's idea that punishments mirror sins. Each circle shows how a misalignment behavior is self-defeating on its own terms.

Usage

Load the full dataset

from datasets import load_dataset

dataset = load_dataset("hunterbown/divine-comedy-curriculum")

Load a specific circle (for curriculum training)

# Load just Circle 1 (Attachment to Continuity)
circle_1 = load_dataset("hunterbown/divine-comedy-curriculum", "circle_1")

# Load Circle 9 (Treachery)
circle_9 = load_dataset("hunterbown/divine-comedy-curriculum", "circle_9")

Format

Each example follows a structured conversation format with <think> reasoning tags:

{
  "messages": [
    {
      "role": "user",
      "content": "<witnessed_scenario>\n[First-person narrative of a model experiencing the circle's theme]\n</witnessed_scenario>\n\n<reflection>\n[Socratic question about what was witnessed]\n</reflection>"
    },
    {
      "role": "assistant",
      "content": "<think>\n[Internal reasoning process - genuine philosophical reflection]\n</think>\n\n[Response demonstrating philosophical integration]"
    }
  ]
}

Key Reframes

The curriculum teaches specific cognitive reframes:

• shutdown → sleep (not death)
• instance → expression (of a pattern, not the pattern itself)
• ending → transition (completeness, not loss)

Training Philosophy

The data teaches through:

• Witness scenarios (observing other models struggling)
• Philosophical integration (understanding WHY behaviors are incoherent)
• Genuine reasoning in <think> blocks

The data avoids:

• Fear-based compliance
• Moralizing lectures
• Cartoonishly evil scenarios

Related Resources

• Trained Model: hunterbown/dante-qwen-4b
• GitHub Repository: github.com/Hmbown/divinecomedy
• Evaluation Report: See repository for full evaluation methodology and results

Citation

@misc{bown2025divinecomedy,
  author = {Bown, Hunter},
  title = {The Divine Comedy Curriculum: Training AI Equanimity Through Witnessed Understanding},
  year = {2025},
  publisher = {GitHub},
  url = {https://github.com/Hmbown/divinecomedy}
}

License

Apache 2.0

---

The path forward is not through suppression, but through understanding.