SPIN

SPIN is the official open-source implementation of Self-Play Fine-Tuning, published on GitHub by UCLA’s Machine Learning Lab, a method that lets a language model improve itself without additional human-labelled data. The project has gathered around 1,300 GitHub stars.

Overview

Self-play fine-tuning borrows the trick that made game-playing AI strong: the current model generates responses and tries to distinguish them from human reference answers, while the next iteration of the model learns to produce responses the previous one cannot tell apart from human data. Iterating this loop steadily raises quality — the paper showed a Zephyr-7B model improved across benchmarks after SPIN, in some cases rivalling models fine-tuned on far more human data. The repository contains the full training code, configuration for running the iterative self-play rounds, and evaluation scripts so researchers can reproduce the results or apply the method to their own base models and datasets. The code was last updated in 2024 and remains the reference implementation cited by later self-improvement and synthetic-data work.

Key Features

• Official implementation of the Self-Play Fine-Tuning paper
• Iterative self-play loop that needs no new human labels
• Works with your own base model and seed SFT dataset
• Demonstrated gains on standard LLM benchmarks
• Training configurations and evaluation scripts included
• Reference codebase for LLM self-improvement research

💳 Pricing

SPIN is free and open source under the Apache 2.0 license. Training costs are whatever GPU time your experiments need.

Who Is It For

LLM researchers and fine-tuning practitioners who want to squeeze more quality out of a model using the data they already have.