Ouroboros
PROJECT
Fiscal Host: Ethicalabs.ai
Ouroboros iteratively refines LLM outputs using local models applying Generate → Critique → Refine loops for structured self-improvement.
About
Ouroboros explores self-refining LLM loops to improve responses iteratively—running entirely on-device with Small Language Models via Ollama.
Initially focused on humanities and philosophy, the project can be applied to math and coding, revealing potential for emergent structured reasoning.
Initially focused on humanities and philosophy, the project can be applied to math and coding, revealing potential for emergent structured reasoning.
The process follows a Generate → Critique → Refine loop, where one LLM generates responses, a smaller model critiques them, and refinements happen based on structured feedback.
Running on an M1 MacBook Air (8GB RAM), the first dataset (~125 samples) took ~8 hours to generate, proving that big models aren’t necessary for iterative improvement on simple tasks.
Future goals include scaling to 100K+ samples, GRPO fine-tuning of SLMs, and federated learning experiments.
Running on an M1 MacBook Air (8GB RAM), the first dataset (~125 samples) took ~8 hours to generate, proving that big models aren’t necessary for iterative improvement on simple tasks.
Future goals include scaling to 100K+ samples, GRPO fine-tuning of SLMs, and federated learning experiments.
Ouroboros isn’t about AGI—it’s a non-agentic experiment in structured LLM refinement.
If you're into self-improving AI and synthetic datasets, check it out on GitHub and Hugging Face! 🚀
If you're into self-improving AI and synthetic datasets, check it out on GitHub and Hugging Face! 🚀
Our team
Massimo Rober...
Admin
Contribute
Become a financial contributor.