Questions on Ouroboros

[TomLucidor]

After seeing Ouro LLM being promoted, it immediately made me want to try and find repos about it... No dice.
But hey, this repo popped up, and I got a lot of questions

1. How can ouroboros prevent catastrophic forgetting when fine-tuning? (RL Razor)
2. Are there ways of avoiding model collapse? (Data Accumulation)
3. Will diffusion or dLLMs be included in the process? They seem very robust
4. Will agentic scaffolds be included? Can they self-improve too? (SEAL + DGM)

[mrs83]

Hi! Just to clarify: this repo is not related to the Ouro LLM paper.

Ouroboros is a tool specifically for Inference and Synthetic Data Generation, not a training framework.

To answer your questions in that context:

• Preventing Model Collapse: We focus on generating high-diversity, high-quality synthetic data. You avoid collapse by mixing this data with real data (Data Accumulation) during your separate training step.
• RL Razor / Forgetting: These are training techniques. This repo provides the data for those training runs, but doesn't handle the weight updates itself.
• Agents/Scaffolds: We are experimenting agentic workflows (and multi-step reasoning) during inference to ensure the synthetic data is high-quality enough to actually improve a model

[TomLucidor]

If you can have everything flushed out and get a wiki/blog going, that would be great! Cus I can't want to see 30B MoE models to get progressively better through fine-tuning to think about complex codebases, software best practices etc.

[mrs83]

That would be interesting and this project is open, anyone can try this with any OpenAI-compatible endpoint.

My work on ethicalabs.ai (a no-profit project that I started in my spare time, while I was still working as employee) is now transitioning to a full-time job, and I currently don’t have time/resources for this project until next year.

Ouroboros started as a framework to test “meta-cognition” capabilities of Language Models (to try let a model “think about thinking”) and it has been later abandoned for synthetic datasets generation purposes.

I consider this as a recursive-inference LM playground, with another LM as a judge.

Taking inspiration from the recursive nature of this project I am now working on new model architectures (not transformers)

[TomLucidor]

Metacognition and recursive self-improvement both sound great! If Diffusion-LM and memory layers can also use these, it would be even better! Can't wait