mirror of
https://github.com/zenlm/zen-docs.git
synced 2026-10-30 14:58:13 +00:00
zeekay
54787245f2
- Fumadocs-based documentation similar to Lux consensus - 14 model documentation pages - Papers index with links to PDFs - Zen LM logo and branding - Dark mode support - GitHub Pages deployment workflow
75 lines
5.4 KiB
TeX
75 lines
5.4 KiB
TeX
\relax
|
||
\providecommand\hyper@newdestlabel[2]{}
|
||
\providecommand\HyField@AuxAddToFields[1]{}
|
||
\providecommand\HyField@AuxAddToCoFields[2]{}
|
||
\citation{deepseek2024}
|
||
\citation{qwen2024}
|
||
\@writefile{toc}{\contentsline {section}{\numberline {1}Introduction}{1}{section.1}\protected@file@percent }
|
||
\citation{mikolov2013efficient}
|
||
\citation{radford2021learning}
|
||
\citation{hinton2015distilling}
|
||
\citation{training_free_grpo2024}
|
||
\@writefile{toc}{\contentsline {subsection}{\numberline {1.1}The Alignment Problem}{2}{subsection.1.1}\protected@file@percent }
|
||
\@writefile{toc}{\contentsline {subsection}{\numberline {1.2}Our Contribution}{2}{subsection.1.2}\protected@file@percent }
|
||
\citation{qwen2024}
|
||
\@writefile{toc}{\contentsline {section}{\numberline {2}Background}{3}{section.2}\protected@file@percent }
|
||
\@writefile{toc}{\contentsline {subsection}{\numberline {2.1}Decentralized Semantic Optimization}{3}{subsection.2.1}\protected@file@percent }
|
||
\@writefile{toc}{\contentsline {subsection}{\numberline {2.2}Qwen3-Embedding-8B}{3}{subsection.2.2}\protected@file@percent }
|
||
\@writefile{toc}{\contentsline {section}{\numberline {3}Method}{4}{section.3}\protected@file@percent }
|
||
\@writefile{toc}{\contentsline {subsection}{\numberline {3.1}Architecture}{4}{subsection.3.1}\protected@file@percent }
|
||
\@writefile{loa}{\contentsline {algorithm}{\numberline {1}{\ignorespaces Zen-Reranker Projection Head}}{4}{algorithm.1}\protected@file@percent }
|
||
\@writefile{toc}{\contentsline {subsection}{\numberline {3.2}Three-Stage Training}{5}{subsection.3.2}\protected@file@percent }
|
||
\@writefile{toc}{\contentsline {subsubsection}{\numberline {3.2.1}Stage 1: Projection Expansion}{5}{subsubsection.3.2.1}\protected@file@percent }
|
||
\@writefile{toc}{\contentsline {subsubsection}{\numberline {3.2.2}Stage 2: Reranking Fine-tuning}{5}{subsubsection.3.2.2}\protected@file@percent }
|
||
\@writefile{toc}{\contentsline {subsubsection}{\numberline {3.2.3}Stage 3: DSO Optimization}{6}{subsubsection.3.2.3}\protected@file@percent }
|
||
\@writefile{toc}{\contentsline {subsection}{\numberline {3.3}Total Training Cost}{6}{subsection.3.3}\protected@file@percent }
|
||
\@writefile{lot}{\contentsline {table}{\numberline {1}{\ignorespaces Training cost breakdown (8× H100 at \$25/GPU-hour)}}{7}{table.1}\protected@file@percent }
|
||
\@writefile{toc}{\contentsline {section}{\numberline {4}Experiments}{7}{section.4}\protected@file@percent }
|
||
\@writefile{toc}{\contentsline {subsection}{\numberline {4.1}Experimental Setup}{7}{subsection.4.1}\protected@file@percent }
|
||
\@writefile{toc}{\contentsline {subsection}{\numberline {4.2}MTEB Results}{7}{subsection.4.2}\protected@file@percent }
|
||
\@writefile{lot}{\contentsline {table}{\numberline {2}{\ignorespaces MTEB benchmark results. Zen-Reranker achieves +0.6 points over base model.}}{7}{table.2}\protected@file@percent }
|
||
\@writefile{toc}{\contentsline {subsection}{\numberline {4.3}DSO Retrieval Accuracy}{8}{subsection.4.3}\protected@file@percent }
|
||
\@writefile{lot}{\contentsline {table}{\numberline {3}{\ignorespaces Cross-model retrieval performance. Native dimension eliminates alignment errors.}}{8}{table.3}\protected@file@percent }
|
||
\@writefile{toc}{\contentsline {subsection}{\numberline {4.4}Compression Efficiency}{8}{subsection.4.4}\protected@file@percent }
|
||
\@writefile{lot}{\contentsline {table}{\numberline {4}{\ignorespaces BitDelta compression ratios. Stage 3 training optimizes for low $\Delta e$ variance.}}{8}{table.4}\protected@file@percent }
|
||
\@writefile{toc}{\contentsline {subsection}{\numberline {4.5}Byzantine Robustness}{9}{subsection.4.5}\protected@file@percent }
|
||
\@writefile{lot}{\contentsline {table}{\numberline {5}{\ignorespaces Byzantine robustness. Median aggregation maintains 97\% of clean performance.}}{9}{table.5}\protected@file@percent }
|
||
\@writefile{toc}{\contentsline {section}{\numberline {5}Discussion}{9}{section.5}\protected@file@percent }
|
||
\@writefile{toc}{\contentsline {subsection}{\numberline {5.1}Why Native Dimension Matters}{9}{subsection.5.1}\protected@file@percent }
|
||
\@writefile{toc}{\contentsline {subsection}{\numberline {5.2}Scaling to Other Dimensions}{9}{subsection.5.2}\protected@file@percent }
|
||
\citation{devlin2018bert}
|
||
\citation{reimers2019sentence}
|
||
\citation{wang2022text}
|
||
\citation{xiao2023c}
|
||
\citation{qwen2024}
|
||
\citation{radford2021learning}
|
||
\citation{jia2021scaling}
|
||
\citation{mikolov2013efficient}
|
||
\citation{han2015learning}
|
||
\citation{jacob2018quantization}
|
||
\citation{bitdelta2024}
|
||
\citation{mcmahan2017communication}
|
||
\citation{blanchard2017machine}
|
||
\citation{training_free_grpo2024}
|
||
\@writefile{lot}{\contentsline {table}{\numberline {6}{\ignorespaces Dimension choice analysis. 7680 balances DeepSeek and Qwen compatibility.}}{10}{table.6}\protected@file@percent }
|
||
\@writefile{toc}{\contentsline {subsection}{\numberline {5.3}Future Work}{10}{subsection.5.3}\protected@file@percent }
|
||
\@writefile{toc}{\contentsline {section}{\numberline {6}Related Work}{10}{section.6}\protected@file@percent }
|
||
\@writefile{toc}{\contentsline {section}{\numberline {7}Conclusion}{10}{section.7}\protected@file@percent }
|
||
\bibcite{deepseek2024}{1}
|
||
\bibcite{qwen2024}{2}
|
||
\bibcite{mikolov2013efficient}{3}
|
||
\bibcite{radford2021learning}{4}
|
||
\bibcite{hinton2015distilling}{5}
|
||
\bibcite{training_free_grpo2024}{6}
|
||
\bibcite{devlin2018bert}{7}
|
||
\bibcite{reimers2019sentence}{8}
|
||
\bibcite{wang2022text}{9}
|
||
\bibcite{xiao2023c}{10}
|
||
\bibcite{jia2021scaling}{11}
|
||
\bibcite{han2015learning}{12}
|
||
\bibcite{jacob2018quantization}{13}
|
||
\bibcite{bitdelta2024}{14}
|
||
\bibcite{mcmahan2017communication}{15}
|
||
\bibcite{blanchard2017machine}{16}
|
||
\gdef \@abspage@last{12}
|