FNAS ● research
genetic architecture search for fractal models
FNAS — Fractal Neural Architecture Search — asks a simple question: if I can't hand-tune the perfect fractal model, can I evolve it? It treats a model's architecture as a genome and runs a genetic algorithm over byte-level language models, breeding and mutating candidates across generations.
It closes the whole loop. It can scrape and distill its own data, train candidate genomes, rank them on a leaderboard, and use a local LLM as an automated judge to score quality — then refine its search around the best genome and keep going, continuously, for as many cycles as you let it. V2 dropped the scraping stack to focus on real corpora and tighter automation.
// highlights
- Genetic search over fractal model genomes
- LLM-as-judge automated scoring + leaderboard
- Self-contained data pipeline: scrape → distill → train
- Continuous, resumable multi-cycle search
- Curses TUI + CLI for the whole workflow
// stack
- Python
- genetic algorithms
- Ollama
- byte-level