RandOpt

Why we included this project

RandOpt is the official codebase for the Neural Thickets paper, which proposes a surprisingly simple post-training idea: sample random weight perturbations around a pretrained model, keep the ones that perform well, and ensemble them. The authors show this can hold its own against more established methods like PPO, GRPO, and evolutionary search. The repo gives you everything needed to try it yourself: Slurm scripts for single and multi-node runs, a local mode, Docker images, a guide for custom datasets, and a distillation script that merges the top-K ensemble back into one model. The iterative-randopt branch even plugs into existing verl or Hugging Face TRL pipelines. It's a research artifact, so expect to adapt it to your own setup, but it's a solid starting point for anyone curious about black-box, ensemble-based post-training.