Size classes widen for large models (x1.25 up to 30M -> x1.5 from 100M)

leaderboard.py

@@ -10,9 +10,22 @@ import html
 import math
 
 MIN_GAMES_GLOW = 2  # a leader needs more than one ranked game, so a single lucky win doesn't glow
-# Size classes are relative, not absolute: a model's neighbourhood runs from size/1.25 to size*1.25
-# (-20% to +25%). A fixed ±20M window lumped every model under 20M into one class, so only one tiny model could glow.
-SIZE_RATIO = 1.25
+# Size classes are relative, not absolute. Two models are neighbours when the bigger one is at most `size_ratio`
+# times the smaller one. The ratio is 1.25 (-20%/+25%) up to 30M and widens smoothly (log scale) to 1.5
+# (-33%/+50%) from 100M on, so e.g. 50M vs 65M and 100M vs 135M share a class while tiny models stay fine-grained.
+RATIO_SMALL, RATIO_LARGE = 1.25, 1.5
+RATIO_FROM, RATIO_TO = 30e6, 100e6
+
+
+def size_ratio(p: float) -> float:
+    t = min(1.0, max(0.0, math.log10(p / RATIO_FROM) / math.log10(RATIO_TO / RATIO_FROM)))
+    return RATIO_SMALL + (RATIO_LARGE - RATIO_SMALL) * t
+
+
+def neighbours(a: float, b: float) -> bool:
+    """Symmetric: the allowed ratio is taken at the pair's geometric-mean size."""
+    return max(a, b) / min(a, b) <= size_ratio(math.sqrt(a * b))
+
 
 _CDN = "https://cdn-avatars.huggingface.co/v1/production/uploads/"
 # owner on the Hub -> (display name, logo, "r, g, b", border colour). Same identities as the BananaMind leaderboard.
@@ -93,14 +106,14 @@ def leaders(entries, gap: int = 0) -> set:
     """Size-class leaders: the best Elo among rated models of similar relative size.
 
     A model glows when (1) it has at least MIN_GAMES_GLOW ranked games, (2) its Elo is above the 1000 start,
-    (3) at least one other rated model sits in its neighbourhood [size/SIZE_RATIO, size*SIZE_RATIO], and
+    (3) at least one other rated model is a size neighbour (see `neighbours`), and
     (4) none of those neighbours has a higher Elo. Ties glow together. `gap` is kept for API compatibility."""
     out = set()
     for e in entries:
         p = e.get("params")
         if not p or e.get("games", 0) < MIN_GAMES_GLOW or e["elo"] <= 1000:
             continue
-        peers = [o for o in entries if o.get("params") and p / SIZE_RATIO <= o["params"] <= p * SIZE_RATIO]
+        peers = [o for o in entries if o.get("params") and neighbours(p, o["params"])]
         if len(peers) < 2:
             continue  # no neighbour to compare with
         if all(o["elo"] <= e["elo"] for o in peers):
@@ -270,8 +283,9 @@ def leaderboard_html(entries, protocol: str, gap: int, gap_label: str) -> str:
     return (
         '<div class="lb-root">'
         f'<section class="lb-panel lb-chart-panel">{views}'
-        f'<div class="lb-footer"><span><i class="lb-rainbow-dot"></i>Size-class leader: best Elo among models from −20% to +25% of its size '
-        f'(min. {MIN_GAMES_GLOW} ranked games, Elo above 1000)</span><span>{len(entries)} models · click a model to open it ↗</span></div></section>'
+        f'<div class="lb-footer"><span><i class="lb-rainbow-dot"></i>Size-class leader: best Elo among models of similar size '
+        f'(−20%/+25% up to 30M, widening to −33%/+50% from 100M · min. {MIN_GAMES_GLOW} ranked games, Elo above 1000)</span>'
+        f'<span>{len(entries)} models · click a model to open it ↗</span></div></section>'
         '<section class="lb-panel lb-param-panel"><div class="lb-heading"><div><h2>Elo vs. Parameters</h2>'
         '<p class="lb-muted">Model size on a logarithmic scale · higher and further left is better</p></div></div>'
         '<div class="lb-key"><span><i class="lb-quadrant-key"></i>Fewer parameters, higher Elo</span><span>┈ Pareto line</span></div>'