Default each family's chart to its best-improvement metric

A configured chart default_metric (family or global) still wins, and an
explicitly picked metric (button click or #metric= link) is pinned across
family switches. Strip the stale ips defaults from the CV families so the
automatic pick applies there.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01AGJFdqmf4AXcsYvb56LLCC

app.js

@@ -410,6 +410,11 @@ const filters = { baseFamily: BASE_FAMILY_KEYS[0] || "", variant: null };
 config.filters.forEach(f => { filters[f.column] = ""; });
 filters.metric = CHART_CFG.default_metric || config.metrics[0]?.column || "";
 
+// [speedup] Whether the current metric was chosen explicitly (metric button
+// click or #metric= in a shared link). While unpinned, each family switch
+// re-picks the metric with the best improvement over the baseline.
+let metricPinned = false;
+
 // ─── Render button groups ─────────────────────────────────────────────────────
 
 function renderBtnGroup(container, items, activeValue) {
@@ -489,6 +494,7 @@ function handleFilterClick(e) {
     btn.classList.add("active");
     const key = group.id.replace("filter-", "");
     filters[key] = btn.dataset.value;
+    if (key === "metric") metricPinned = true;   // [speedup] user's explicit pick
     render();
 }
 
@@ -971,6 +977,7 @@ const urlstate = initUrlState({
     config, filters, BASE_FAMILIES, deriveBaseFamily, GROUP_BY,
     availableOptions, renderSidebar, updateDependentFilters,
     switchBaseFamily, render,
+    onExplicitMetric: () => { metricPinned = true; },  // [speedup] linked metric wins
 });
 
 // ─── Embed SVG module ────────────────────────────────────────────────────────
@@ -1043,11 +1050,21 @@ async function switchBaseFamily(baseFamilyKey) {
     // Fall back to a metric that has data in this family (e.g. LLM tps vs CV ips)
     const hasData = (col) => DATA.some(r =>
         r[col] !== undefined && r[col] !== null && r[col] !== 0 && r[col] !== "N/A");
+    // [speedup] Default the chart metric, unless one was picked explicitly
+    // (and still has data here). A configured default_metric — the family's
+    // own chart config or the global one, same resolution buildChart uses —
+    // overrides the automatic best-improvement-over-baseline pick.
+    if (!hasData(filters.metric)) metricPinned = false;
+    if (!metricPinned) {
+        const chartCfg = config.model_families?.[activeFamilyKey()]?.chart || CHART_CFG;
+        const cfgDefault = chartCfg.default_metric;
+        const pick = (cfgDefault && hasData(cfgDefault))
+            ? cfgDefault
+            : speedup.bestMetric(DATA, metricsWithData(DATA));
+        if (pick) filters.metric = pick;
+    }
     if (!hasData(filters.metric)) {
-        const familyDefault = config.model_families?.[activeFamilyKey()]?.chart?.default_metric;
-        filters.metric = (familyDefault && hasData(familyDefault))
-            ? familyDefault
-            : (config.metrics.find(m => hasData(m.column))?.column || filters.metric);
+        filters.metric = config.metrics.find(m => hasData(m.column))?.column || filters.metric;
     }
     render();
 }

config.json

@@ -306,7 +306,6 @@
       "data_file": "data/DINOv3-ViT-B16.csv",
       "accuracy_url": "https://huggingface.co/embedl/dinov3-quantized-tensorrt",
       "chart": {
-        "default_metric": "ips",
         "scenarios": []
       },
       "default_device": "agx_orin",
@@ -348,7 +347,6 @@
       "data_file": "data/SAM-3D-Body.csv",
       "accuracy_url": "https://huggingface.co/embedl/sam-3d-body",
       "chart": {
-        "default_metric": "ips",
         "scenarios": []
       },
       "default_device": "l4",
@@ -363,7 +361,6 @@
       "data_file": "data/Parakeet-TDT-0.6B-v3.csv",
       "accuracy_url": "https://huggingface.co/embedl/parakeet-tdt-0.6b-v3-quantized-tensorrt",
       "chart": {
-        "default_metric": "ips",
         "scenarios": []
       },
       "accuracy_higher_is_better": false,
@@ -379,7 +376,6 @@
       "data_file": "data/Chronos-2.csv",
       "accuracy_url": "https://huggingface.co/embedl/chronos-2-quantized-trt",
       "chart": {
-        "default_metric": "ips",
         "scenarios": [
           {
             "label": "Context 512",
@@ -407,7 +403,6 @@
       "data_file": "data/MobileViT-Small.csv",
       "accuracy_url": "https://huggingface.co/embedl/mobilevit-small-quantized",
       "chart": {
-        "default_metric": "ips",
         "scenarios": []
       },
       "default_device": "agx_orin",
@@ -421,7 +416,6 @@
       "data_file": "data/all-MiniLM-L6-v2.csv",
       "accuracy_url": "https://huggingface.co/embedl/all-MiniLM-L6-v2-quantized-trt",
       "chart": {
-        "default_metric": "ips",
         "scenarios": []
       },
       "default_device": "agx_orin",
@@ -435,7 +429,6 @@
       "data_file": "data/paraphrase-multilingual-MiniLM-L12-v2.csv",
       "accuracy_url": "https://huggingface.co/embedl/paraphrase-multilingual-MiniLM-L12-v2-quantized-trt",
       "chart": {
-        "default_metric": "ips",
         "scenarios": []
       },
       "default_device": "agx_orin",
@@ -449,7 +442,6 @@
       "data_file": "data/RF-DETR-Base.csv",
       "accuracy_url": "https://huggingface.co/embedl/rf-detr-base-quantized-tensorrt",
       "chart": {
-        "default_metric": "ips",
         "scenarios": []
       },
       "default_device": "orin_nano_super",

speedup.js

@@ -1,7 +1,7 @@
 // ─── Speedup vs Baseline Column ──────────────────────────────────────────────
 //
 // Extracted module. Call initSpeedup(deps) from the main app once globals are
-// ready.  Returns { prepare, headerHtml, cellHtml, metricCol }.
+// ready.  Returns { prepare, headerHtml, cellHtml, metricCol, bestMetric }.
 //
 // The benchmark data is inherently paired: an external baseline model (e.g.
 // meta-llama/Llama-3.2-1B-Instruct) and the optimized embedl variants of it
@@ -45,6 +45,50 @@ function formatRatio(ratio) {
 // Per-table state, rebuilt by prepare() before each table is rendered.
 let state = null;
 
+// ─── Best metric ─────────────────────────────────────────────────────────────
+
+/**
+ * The metric whose optimized-vs-baseline improvement is largest across `rows`,
+ * so the chart can lead with the model's strongest result (highest speedup,
+ * biggest memory/latency reduction, …). Ratios are oriented so that > 1 always
+ * means "better than baseline", making metrics directly comparable.
+ * @param rows     all rows of the loaded family
+ * @param metrics  candidate metric configs (usually the ones with data)
+ * @returns the winning metric column, or null when nothing pairs
+ */
+function bestMetric(rows, metrics) {
+    if (!ENABLED || !rows || !rows.length) return null;
+
+    const baselines = {};
+    rows.forEach(r => {
+        if (!isExternalModel(r[MODEL_COL])) return;
+        const k = pairKey(r, KEY_COLS);
+        (baselines[k] = baselines[k] || []).push(r);
+    });
+
+    let best = null;
+    (metrics || []).forEach(m => {
+        const hib = m.higher_is_better !== false;
+        let top;
+        rows.forEach(r => {
+            if (isExternalModel(r[MODEL_COL])) return;
+            const cands = baselines[pairKey(r, KEY_COLS)];
+            if (!cands || !cands.length) return;
+            if (new Set(cands.map(c => c[MODEL_COL])).size > 1) return;
+            const b = usable(cands[0][m.column]);
+            const o = usable(r[m.column]);
+            if (b === undefined || o === undefined) return;
+            const ratio = hib ? o / b : b / o;
+            if (!isFinite(ratio) || ratio <= 0) return;
+            if (top === undefined || ratio > top) top = ratio;
+        });
+        if (top !== undefined && (!best || top > best.ratio)) {
+            best = { column: m.column, ratio: top };
+        }
+    });
+    return best ? best.column : null;
+}
+
 // ─── Pairing ─────────────────────────────────────────────────────────────────
 
 /**
@@ -128,6 +172,6 @@ function cellHtml(row) {
         + `<span class="speedup-val${down}">${formatRatio(ratio)}</span></td>`;
 }
 
-return { prepare, headerHtml, cellHtml, metricCol };
+return { prepare, headerHtml, cellHtml, metricCol, bestMetric };
 
 }

urlstate.js

@@ -18,7 +18,7 @@ function initUrlState(deps) {
 const {
     config, filters, BASE_FAMILIES, deriveBaseFamily, GROUP_BY,
     availableOptions, renderSidebar, updateDependentFilters,
-    switchBaseFamily, render,
+    switchBaseFamily, render, onExplicitMetric,
 } = deps;
 
 const ENABLED     = config.url_state !== false;
@@ -92,6 +92,7 @@ function applyParams(params) {
 
     if (params.metric && METRIC_COLS.includes(params.metric)) {
         filters.metric = params.metric;
+        onExplicitMetric?.();   // a linked metric beats the best-improvement default
     }
 
     pending = {};