Training pipeline (with loss/optimizer loop)

\documentclass[border=6pt]{standalone}

% --- packages (mirror these in template.meta.json "requires") ---
\usepackage{tikz}
\usetikzlibrary{positioning, arrows.meta}

% --- palette (canonical source: reference/color-palettes/color-palettes.md; light variant) ---
\definecolor{otblue}{HTML}{0072B2}
\definecolor{otorange}{HTML}{E69F00}
\definecolor{otteal}{HTML}{009E73}
\definecolor{otpurple}{HTML}{CC79A7}
\definecolor{otgray}{HTML}{5A5A5A}

\begin{document}
% ==== parameters (edit these) ============================================
\def\nodesep{1.2}   % horizontal gap between forward stages (cm)
\def\loopdrop{1.6}  % how far below the model the optimizer sits (cm)
\def\boxminw{1.9}   % minimum stage width (cm)
\def\boxminh{1.0}   % minimum stage height (cm)
% labels
\def\datalabel{Dataset}
\def\modellabel{Model}
\def\losslabel{Loss}
\def\optlabel{Optimizer}
\def\labelslabel{Labels $y$}
\def\predlabel{$\hat{y}$}
\def\gradlabel{gradients $\nabla\mathcal{L}$}
\def\updatelabel{update $\theta$}
% =========================================================================

\begin{tikzpicture}[
    node distance=\nodesep cm,
    >={Stealth[length=2.4mm]},
    stage/.style={
      draw=otgray!55, line width=0.7pt, rounded corners=3pt, align=center,
      minimum width=\boxminw cm, minimum height=\boxminh cm, font=\sffamily\small,
    },
    io/.style={stage, fill=otgray!12},
    fwd/.style={draw=otgray!70, line width=0.9pt, ->},
    back/.style={draw=otpurple!85!black, line width=1.0pt, ->},
    elabel/.style={font=\sffamily\footnotesize, text=otgray},
    blabel/.style={font=\sffamily\footnotesize, text=otpurple!85!black},
  ]

  % --- forward path: Dataset -> Model -> Loss, with Labels feeding the Loss ---
  \node[io] (data) {\datalabel};
  \node[stage, fill=otblue!18, right=of data] (model) {\modellabel};
  \node[stage, fill=otorange!20, right=of model] (loss) {\losslabel};
  \node[io, above=of loss] (labels) {\labelslabel};

  % --- optimizer sits below the model on the return path ---
  \node[stage, fill=otteal!18, below=\loopdrop cm of model] (opt) {\optlabel};

  % forward arrows
  \draw[fwd] (data) -- (model);
  \draw[fwd] (model) -- (loss) node[midway, above, elabel] {\predlabel};
  \draw[fwd] (labels) -- (loss);

  % --- training loop: backprop gradients then the optimizer update ---
  \draw[back] (loss.south) |- (opt.east)
    node[pos=0.65, above, blabel] {\gradlabel};
  \draw[back] (opt.north) -- (model.south)
    node[midway, right, blabel] {\updatelabel};

\end{tikzpicture}
\end{document}