Files
papers/zip-002-zen-reranker/sections/11-comparison-with-alternatives.tex

52 lines
1.4 KiB
TeX

\section{Comparison with Alternatives}
\subsection{BAAI bge-reranker-v2-m3}
\begin{table}[h]
\centering
\caption{Comparison with BAAI bge-reranker-v2-m3}
\begin{tabular}{lrr}
\toprule
\textbf{Metric} & \textbf{bge-reranker-v2-m3} & \textbf{Zen-Reranker-8B} \\
\midrule
Dimension & 1024 & 7680 \\
Parameters & 568M & 8.3B \\
MTEB Avg & 64.7 & 68.4 \\
MTEB Retrieval & 58.1 & 62.7 \\
DSO Recall@5 & 81.3\% & 94.7\% \\
Compression Ratio & 26.9$\times$ & 31.87$\times$ \\
Latency (ms) & 16.2 & 21.5 \\
Training Cost & \$8,000 & \$10,800 \\
\bottomrule
\end{tabular}
\end{table}
\textbf{Trade-offs:}
\begin{itemize}
\item BGE: Faster inference, lower cost
\item \textbf{Zen-Reranker: Higher accuracy, better DSO performance, optimal compression}
\end{itemize}
\subsection{jina-embeddings-v3}
\begin{table}[h]
\centering
\caption{Comparison with jina-embeddings-v3}
\begin{tabular}{lrr}
\toprule
\textbf{Metric} & \textbf{jina-embeddings-v3} & \textbf{Zen-Reranker-8B} \\
\midrule
Dimension (configurable) & 768-8192 & 7680 (native) \\
Parameters & 570M & 8.3B \\
MTEB Avg & 66.2 & 68.4 \\
DSO Cross-Model & 85.7\% & 94.7\% \\
Byzantine Robustness & Not evaluated & 92.1\% @ 30\% attack \\
License & Apache 2.0 & Apache 2.0 \\
\bottomrule
\end{tabular}
\end{table}
\textbf{Key Difference:} Jina supports variable dimensions via truncation/padding, but not optimized for 7680-dim DSO. Zen-Reranker is purpose-built.