mirror of
https://github.com/zoo-apps/papers.git
synced 2026-10-30 14:56:54 +00:00
zeekayandHanzo Dev
dced1dcfb9
Bylines, paper authorship, correspondence addresses and the founder blurb under each post named an individual. They name the organisation that published the work instead: the brand on posts and pages, the research byline on papers, and research@ in place of a personal address. Both are what these repositories already used elsewhere — nothing new was coined. Co-authors are untouched. Where a paper named someone else alongside, they stay named; only the one byline changed. Bibliography entries, running heads, PDF metadata and affiliations move with it, so a citation still resolves to whoever the cited paper names. Sample and fixture data uses the placeholder name these tests already use elsewhere, so no assertion changed meaning. Co-authored-by: Hanzo Dev <[email protected]>
52 lines
2.7 KiB
TeX
52 lines
2.7 KiB
TeX
\documentclass[11pt,letterpaper]{article}
|
|
\usepackage{shared/zoocover}
|
|
\usepackage[utf8]{inputenc}
|
|
\usepackage[T1]{fontenc}
|
|
\usepackage{hyperref}
|
|
\usepackage{url}
|
|
\usepackage{booktabs}
|
|
\usepackage{amsfonts}
|
|
\usepackage{amsmath}
|
|
\usepackage{amssymb}
|
|
\usepackage{nicefrac}
|
|
\usepackage{microtype}
|
|
\usepackage{graphicx}
|
|
\usepackage{natbib}
|
|
\usepackage{doi}
|
|
% geometry loaded by zoocover
|
|
|
|
\title{Gym: Democratizing AI Model Training Through\\Comprehensive Open-Source Infrastructure}
|
|
|
|
\author{
|
|
Antje Worring, Zoo Labs Foundation Research\thanks{[email protected]}\\
|
|
\textit{Hanzo Industries \quad Lux Industries \quad Zoo Labs Foundation}\\
|
|
\texttt{[email protected]}
|
|
}
|
|
|
|
\date{September 2025}
|
|
|
|
\begin{document}
|
|
|
|
\zoocoverpage
|
|
|
|
\begin{abstract}
|
|
The democratization of artificial intelligence remains hindered by the substantial computational costs and technical complexity of model training. We present \textbf{Gym}, an open-source AI model training platform designed to lower these barriers through comprehensive infrastructure supporting 100+ model architectures, efficient training algorithms, and accessible interfaces. Originally forked from LLaMA Factory in May 2023, Gym has evolved over 2.5 years into a production-ready system incorporating state-of-the-art methods including full fine-tuning, Low-Rank Adaptation (LoRA), Reinforcement Learning from Human Feedback (RLHF), and recently Group Relative Policy Optimization (GRPO). Our Training-Free GRPO implementation achieves comparable performance to gradient-based methods at 99.8\% cost reduction (\$18 vs \$10,000+ for domain adaptation), demonstrating that democratization need not compromise capability. Gym has enabled researchers worldwide to fine-tune models ranging from 0.5B to 72B parameters on consumer hardware, facilitated educational initiatives across 50+ institutions, and powered production deployments serving millions of users. As a 501(c)(3) non-profit project, Gym embodies the principle that advanced AI should be accessible to all, not just well-resourced laboratories. Code, documentation, and pre-trained models are available at \url{https://github.com/zooai/gym}.
|
|
\end{abstract}
|
|
|
|
\input{sections/01-introduction}
|
|
\input{sections/02-background-and-motivation}
|
|
\input{sections/03-architecture-overview}
|
|
\input{sections/04-training-methods}
|
|
\input{sections/05-multi-modal-training}
|
|
\input{sections/06-performance-optimizations}
|
|
\input{sections/07-distributed-training}
|
|
\input{sections/08-educational-impact}
|
|
\input{sections/09-integration-with-zoo-ecosystem}
|
|
\input{sections/10-evaluation-and-benchmarks}
|
|
\input{sections/11-comparison-with-related-work}
|
|
\input{sections/12-future-work}
|
|
\input{sections/13-conclusion}
|
|
\input{sections/99-acknowledgments}
|
|
|
|
\end{document}
|