Fix Specialists

Fix Specialists and Half Specialists by classifying a model as specialist if the Arithmark 2 score is more than 30% higher in points than Hellaswag. This fixes models like Atom 2.7 While Also Fixing models like Nexus Erebus 135M that train on alot of synthetic arithmetic and almost no actual web corpus that didn't get flagged as specialist before. For models that actually perform like this there is a exclusion list.

index.html

@@ -57,14 +57,14 @@
       font-weight: 400;
       letter-spacing: -0.2px;
     }
-    
+
     .hero-sub a {
       color: var(--accent);
       text-decoration: none;
       border-bottom: 1px solid transparent;
       transition: border-color 0.15s;
     }
-    
+
     .hero-sub a:hover {
       border-bottom-color: var(--accent);
     }
@@ -929,12 +929,12 @@
         To support our work and help us keep this leaderboard up to date, please consider giving the space a like and follow!
       </p>
     </section>
-    
+
     <!-- HIGHLIGHTS -->
     <section id="highlights" style="padding-top:20px;">
       <div class="insight-grid" id="insight-grid"></div>
     </section>
-    
+
     <!-- NOTABLE NEW RELEASES -->
     <section id="new-releases" style="padding-top:40px;">
       <div class="release-head">
@@ -1248,12 +1248,12 @@
     { name: 'BananaMind-2-Medium-Chat', org: 'bananamind', params: 49559552, paramsDisplay: '49.6M', arc: 43.43, hellaswag: 31.71, piqa: 60.66, arcChall: 24.40, arithmark2: 29.92, links: { card: 'https://huggingface.co/BananaMind/BananaMind-2-Medium-Chat' } },
     { name: 'Photon-2.0-1M',        org: 'atomixlabs', params: 1049728, paramsDisplay: '1.0M', arc: 30.05, hellaswag: 28.17, piqa: 53.92, arcChall: 22.53, arithmark2: 26.64, links: { card: 'https://huggingface.co/AtomixLabs/Photon-2.0-1M' } },
     { name: 'Nexus-Erebus-135M', org: 'ideoalabs', params: 134000000, paramsDisplay: '135M', arc: 48.61, hellaswag: 30.61, piqa: 64.53, arcChall: 25.43, arithmark2: 66.68, links: { card: 'https://huggingface.co/MaliosDark/Nexus-Erebus-135M' } },
-  ];  
+  ];
 
   // ═══════════════════════════════════════════════════════════════
   // ═══ STATE
   // ═══════════════════════════════════════════════════════════════
-  
+
   const BENCHMARKS = ['arc', 'hellaswag', 'piqa', 'arcChall', 'arithmark2'];
   const METRICS = [
     { key: 'avg', label: 'Avg', fullLabel: 'Average Score' },
@@ -1277,8 +1277,16 @@
   let activeBenchmark = 'avg';
   let chartInstances = {};
   const ORG_MOVEMENT_CUTOFF_TYPE = 'orgMovementCutoff';
-  const SPECIALIST_MIN_FIELD_Z = 3.5;
-  const SPECIALIST_MIN_Z_GAP_TO_NEXT_METRIC = 2.5;
+  // 'relative' → ArithMark-2 > HellaSwag * 1.30   (30% higher)
+  // 'points'   → ArithMark-2 - HellaSwag > 30     (30 percentage points higher)
+  const SPECIALIST_MODE = 'points';
+  const SPECIALIST_MARGIN = 30;
+  // Models that must never be flagged as specialists, regardless of the rule.
+  // Names must match the `name` field in MODELS exactly.
+  const SPECIALIST_EXCLUSIONS = new Set([
+  // 'MobileLLM-R1-140M-base',
+  // 'Quark-135M',
+]);
 
   function isBelowOrgMovementCutoff(model) {
     const cutoffIndex = MODELS.findIndex(m => m.type === ORG_MOVEMENT_CUTOFF_TYPE);
@@ -1321,37 +1329,28 @@
     return item ? (full ? item.fullLabel : item.label) : metric;
   }
 
-  function getSpecialistMetric(model, models = getModelEntries(MODELS)) {
-    const zScores = BENCHMARKS.map(metric => {
-      const score = model[metric];
-      if (score === null || score === undefined) return null;
-
-      const fieldScores = getModelEntries(models)
-        .map(m => m[metric])
-        .filter(v => v !== null && v !== undefined);
-      if (fieldScores.length < 2) return null;
-
-      const mean = fieldScores.reduce((sum, v) => sum + v, 0) / fieldScores.length;
-      const std = getStdDev(fieldScores);
-      if (!std) return null;
+  function getSpecialistMetric(model) {
+  if (!model || model.type) return null;
+  if (SPECIALIST_EXCLUSIONS.has(model.name)) return null;
 
-      return { metric, z: (score - mean) / std, score };
-    }).filter(Boolean).sort((a, b) => b.z - a.z);
+  const arith = model.arithmark2;
+  const hella = model.hellaswag;
+  if (arith === null || arith === undefined) return null;
+  if (hella === null || hella === undefined || hella <= 0) return null;
 
-    if (zScores.length < 2) return null;
+  const flagged = SPECIALIST_MODE === 'relative'
+    ? arith > hella * (1 + SPECIALIST_MARGIN / 100)
+    : (arith - hella) > SPECIALIST_MARGIN;
 
-    const [top, next] = zScores;
-    return top.z >= SPECIALIST_MIN_FIELD_Z && (top.z - next.z) >= SPECIALIST_MIN_Z_GAP_TO_NEXT_METRIC
-      ? top.metric
-      : null;
-  }
+  return flagged ? 'arithmark2' : null;
+}
 
-  function isSpecialistModel(model, models = getFilteredModels()) {
-    return Boolean(getSpecialistMetric(model));
-  }
+  function isSpecialistModel(model) {
+  return Boolean(getSpecialistMetric(model));
+}
   function getModelEntries(models = MODELS) {
-    return models.filter(m => !m.type);
-  }
+  return models.filter(m => !m.type);
+}
 
   function getFilteredModels(models = MODELS) {
   let arr = getModelEntries(models);