papers/gym-training-platform/sections/11-comparison-with-related-work.tex

\section{Comparison with Related Work}


\begin{table}[h]
\centering
\small
\begin{tabular}{lcccccc}
\toprule
Feature & Gym & Axolotl & llama-recipes & TRL & Ludwig & AutoTrain \\
\midrule
Web UI & ✓ & ✗ & ✗ & ✗ & ✓ & ✓ \\
CLI & ✓ & ✓ & ✓ & ✗ & ✓ & ✓ \\
Python API & ✓ & ✓ & ✓ & ✓ & ✓ & ✓ \\
Model Count & 112 & 80+ & 10 & Any & Any & 50+ \\
Multi-Modal & ✓ & ✓ & ✗ & ✗ & ✓ & ✓ \\
RLHF Methods & 6 & 3 & 1 & 5 & 0 & 2 \\
Training-Free & ✓ & ✗ & ✗ & ✗ & ✗ & ✗ \\
Quantization & 4/8-bit & 4/8-bit & 8-bit & ✗ & ✗ & 4/8-bit \\
Educational Focus & ✓ & ✗ & ✗ & ✗ & ✓ & ✗ \\
501(c)(3) Status & ✓ & ✗ & ✗ & ✗ & ✗ & ✗ \\
\bottomrule
\end{tabular}
\caption{Feature comparison with major LLM training frameworks.}
\end{table}

\textbf{Key Differentiators}:
\begin{itemize}
\item \textbf{Gym}: Comprehensive, educational, non-profit, Training-Free GRPO
\item \textbf{Axolotl}: Strong YAML configs, active community, but CLI-only
\item \textbf{llama-recipes}: Official Meta recipes, excellent for LLaMA family
\item \textbf{TRL}: Low-level library, requires coding expertise
\item \textbf{Ludwig}: General ML framework, less LLM-optimized
\item \textbf{AutoTrain}: Commercial, proprietary optimizations, vendor lock-in
\end{itemize}