Spaces:
Running
Running
Update index.html
Browse files- index.html +513 -184
index.html
CHANGED
|
@@ -7,19 +7,20 @@
|
|
| 7 |
<script src="https://cdn.jsdelivr.net/npm/chart.js@4"></script>
|
| 8 |
<style>
|
| 9 |
:root {
|
| 10 |
-
--bg: #
|
| 11 |
-
--bg2: #
|
| 12 |
-
--surface:
|
| 13 |
-
--
|
| 14 |
--border: rgba(255,255,255,0.08);
|
| 15 |
-
--
|
| 16 |
-
--text: #
|
| 17 |
-
--muted: #
|
| 18 |
-
--
|
| 19 |
-
--silver: #
|
| 20 |
-
--
|
| 21 |
-
--
|
| 22 |
-
--
|
|
|
|
| 23 |
}
|
| 24 |
|
| 25 |
* { box-sizing: border-box; }
|
|
@@ -31,27 +32,32 @@
|
|
| 31 |
font-family: Inter, ui-sans-serif, system-ui, -apple-system, BlinkMacSystemFont, "Segoe UI", sans-serif;
|
| 32 |
color: var(--text);
|
| 33 |
background:
|
| 34 |
-
radial-gradient(circle at top left, rgba(255,255,255,0.
|
| 35 |
radial-gradient(circle at bottom right, rgba(255,255,255,0.05), transparent 24%),
|
| 36 |
-
linear-gradient(180deg, #
|
| 37 |
}
|
| 38 |
|
| 39 |
.app {
|
| 40 |
-
max-width:
|
| 41 |
margin: 0 auto;
|
| 42 |
display: grid;
|
| 43 |
gap: 18px;
|
| 44 |
}
|
| 45 |
|
| 46 |
-
.hero,
|
| 47 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 48 |
border: 1px solid var(--border);
|
| 49 |
box-shadow: var(--shadow);
|
|
|
|
| 50 |
}
|
| 51 |
|
| 52 |
.hero {
|
| 53 |
-
border-radius: calc(var(--radius) +
|
| 54 |
-
padding:
|
| 55 |
position: relative;
|
| 56 |
overflow: hidden;
|
| 57 |
}
|
|
@@ -60,12 +66,12 @@
|
|
| 60 |
content: "";
|
| 61 |
position: absolute;
|
| 62 |
inset: 0;
|
| 63 |
-
background: linear-gradient(90deg, rgba(255,255,255,0.03), transparent
|
| 64 |
pointer-events: none;
|
| 65 |
}
|
| 66 |
|
| 67 |
.eyebrow {
|
| 68 |
-
color: var(--
|
| 69 |
text-transform: uppercase;
|
| 70 |
letter-spacing: 0.22em;
|
| 71 |
font-size: 11px;
|
|
@@ -74,24 +80,28 @@
|
|
| 74 |
|
| 75 |
h1 {
|
| 76 |
margin: 0;
|
| 77 |
-
font-size: clamp(34px, 5vw,
|
| 78 |
line-height: 0.96;
|
| 79 |
letter-spacing: -0.06em;
|
| 80 |
font-weight: 850;
|
| 81 |
}
|
| 82 |
|
| 83 |
-
.title-silver {
|
|
|
|
|
|
|
|
|
|
|
|
|
| 84 |
.sub {
|
| 85 |
margin: 14px 0 0;
|
| 86 |
color: var(--muted);
|
| 87 |
-
max-width:
|
| 88 |
line-height: 1.7;
|
| 89 |
font-size: 14px;
|
| 90 |
}
|
| 91 |
|
| 92 |
.layout {
|
| 93 |
display: grid;
|
| 94 |
-
grid-template-columns:
|
| 95 |
gap: 18px;
|
| 96 |
align-items: start;
|
| 97 |
}
|
|
@@ -103,7 +113,10 @@
|
|
| 103 |
top: 18px;
|
| 104 |
}
|
| 105 |
|
| 106 |
-
.panel h2,
|
|
|
|
|
|
|
|
|
|
| 107 |
margin: 0 0 14px;
|
| 108 |
font-size: 14px;
|
| 109 |
letter-spacing: 0.12em;
|
|
@@ -117,7 +130,7 @@
|
|
| 117 |
.section-label {
|
| 118 |
display: block;
|
| 119 |
margin-bottom: 10px;
|
| 120 |
-
color: var(--
|
| 121 |
font-size: 12px;
|
| 122 |
text-transform: uppercase;
|
| 123 |
letter-spacing: 0.12em;
|
|
@@ -129,7 +142,8 @@
|
|
| 129 |
gap: 8px;
|
| 130 |
}
|
| 131 |
|
| 132 |
-
button
|
|
|
|
| 133 |
border: 1px solid rgba(255,255,255,0.10);
|
| 134 |
background: rgba(255,255,255,0.03);
|
| 135 |
color: var(--muted);
|
|
@@ -139,18 +153,21 @@
|
|
| 139 |
transition: all 0.15s ease;
|
| 140 |
font: inherit;
|
| 141 |
font-size: 13px;
|
|
|
|
| 142 |
}
|
| 143 |
|
| 144 |
-
button:hover
|
|
|
|
| 145 |
border-color: rgba(255,255,255,0.24);
|
| 146 |
color: var(--text);
|
| 147 |
transform: translateY(-1px);
|
| 148 |
}
|
| 149 |
|
| 150 |
-
button.active
|
|
|
|
| 151 |
color: #fafafa;
|
| 152 |
-
border-color: rgba(
|
| 153 |
-
background: linear-gradient(180deg, rgba(
|
| 154 |
box-shadow: 0 0 0 1px rgba(255,255,255,0.03) inset;
|
| 155 |
}
|
| 156 |
|
|
@@ -169,7 +186,7 @@
|
|
| 169 |
}
|
| 170 |
|
| 171 |
.stat .k {
|
| 172 |
-
color: var(--
|
| 173 |
font-size: 11px;
|
| 174 |
text-transform: uppercase;
|
| 175 |
letter-spacing: 0.12em;
|
|
@@ -177,14 +194,14 @@
|
|
| 177 |
|
| 178 |
.stat .v {
|
| 179 |
margin-top: 6px;
|
| 180 |
-
font-size:
|
| 181 |
-
font-weight:
|
| 182 |
color: var(--silver);
|
| 183 |
font-variant-numeric: tabular-nums;
|
| 184 |
}
|
| 185 |
|
| 186 |
.small-note {
|
| 187 |
-
color: var(--
|
| 188 |
font-size: 12px;
|
| 189 |
line-height: 1.6;
|
| 190 |
margin-top: 10px;
|
|
@@ -192,13 +209,15 @@
|
|
| 192 |
|
| 193 |
.chart-card {
|
| 194 |
border-radius: var(--radius);
|
| 195 |
-
padding:
|
| 196 |
-
min-height:
|
|
|
|
|
|
|
| 197 |
}
|
| 198 |
|
| 199 |
.chart-top {
|
| 200 |
display: flex;
|
| 201 |
-
align-items:
|
| 202 |
justify-content: space-between;
|
| 203 |
gap: 12px;
|
| 204 |
margin-bottom: 12px;
|
|
@@ -207,8 +226,8 @@
|
|
| 207 |
|
| 208 |
.chart-title {
|
| 209 |
margin: 0;
|
| 210 |
-
font-size:
|
| 211 |
-
font-weight:
|
| 212 |
letter-spacing: -0.03em;
|
| 213 |
}
|
| 214 |
|
|
@@ -216,68 +235,219 @@
|
|
| 216 |
margin: 4px 0 0;
|
| 217 |
color: var(--muted);
|
| 218 |
font-size: 13px;
|
|
|
|
| 219 |
}
|
| 220 |
|
| 221 |
.badge {
|
| 222 |
-
padding:
|
| 223 |
border-radius: 999px;
|
| 224 |
-
background: rgba(
|
| 225 |
-
border: 1px solid rgba(
|
| 226 |
color: var(--silver);
|
| 227 |
font-size: 13px;
|
| 228 |
white-space: nowrap;
|
|
|
|
| 229 |
}
|
| 230 |
|
| 231 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 232 |
|
| 233 |
.list-card {
|
| 234 |
border-radius: var(--radius);
|
| 235 |
padding: 16px;
|
| 236 |
}
|
| 237 |
|
| 238 |
-
.
|
| 239 |
-
display: grid;
|
| 240 |
-
gap: 10px;
|
| 241 |
-
max-height: 360px;
|
| 242 |
overflow: auto;
|
| 243 |
-
|
|
|
|
|
|
|
| 244 |
}
|
| 245 |
|
| 246 |
-
|
| 247 |
-
|
| 248 |
-
|
| 249 |
-
|
| 250 |
-
|
| 251 |
-
|
| 252 |
-
|
| 253 |
-
|
| 254 |
-
|
| 255 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 256 |
cursor: pointer;
|
| 257 |
}
|
| 258 |
|
| 259 |
-
|
| 260 |
-
|
| 261 |
-
border-color: rgba(221,224,230,0.20);
|
| 262 |
-
background: rgba(221,224,230,0.045);
|
| 263 |
}
|
| 264 |
|
| 265 |
-
|
| 266 |
-
|
| 267 |
-
|
|
|
|
|
|
|
|
|
|
| 268 |
|
| 269 |
-
|
| 270 |
-
|
| 271 |
-
.panel { position: static; }
|
| 272 |
-
canvas { height: 500px !important; }
|
| 273 |
}
|
| 274 |
|
| 275 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 276 |
body { padding: 16px; }
|
| 277 |
-
.hero, .panel, .chart-card, .list-card { border-radius: 18px; }
|
| 278 |
.stat-grid { grid-template-columns: 1fr; }
|
| 279 |
-
.model-row { grid-template-columns: 1fr; gap: 8px; }
|
| 280 |
canvas { height: 420px !important; }
|
|
|
|
| 281 |
}
|
| 282 |
</style>
|
| 283 |
</head>
|
|
@@ -287,8 +457,8 @@
|
|
| 287 |
<div class="eyebrow">Regression explorer</div>
|
| 288 |
<h1><span class="title-silver">Silver</span> benchmark fit line</h1>
|
| 289 |
<p class="sub">
|
| 290 |
-
Pick a benchmark, fit a linear regression on <b>log10(parameters)</b>, and inspect
|
| 291 |
-
Hover any dot to see the model name
|
| 292 |
</p>
|
| 293 |
</section>
|
| 294 |
|
|
@@ -301,16 +471,29 @@
|
|
| 301 |
<div class="btn-grid" id="benchmarkButtons"></div>
|
| 302 |
</div>
|
| 303 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 304 |
<div class="section">
|
| 305 |
<span class="section-label">Regression stats</span>
|
| 306 |
<div class="stat-grid">
|
| 307 |
<div class="stat"><div class="k">Slope</div><div class="v" id="statSlope">—</div></div>
|
| 308 |
<div class="stat"><div class="k">Intercept</div><div class="v" id="statIntercept">—</div></div>
|
| 309 |
-
<div class="stat"><div class="k">MSE</div><div class="v" id="
|
| 310 |
-
<div class="stat"><div class="k">RMSE</div><div class="v" id="
|
| 311 |
</div>
|
| 312 |
<div class="small-note" id="fitNote">The fit updates instantly when you switch benchmarks.</div>
|
| 313 |
</div>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 314 |
</aside>
|
| 315 |
|
| 316 |
<main style="display:grid; gap:18px;">
|
|
@@ -319,64 +502,147 @@
|
|
| 319 |
<div>
|
| 320 |
<h2 class="chart-title" id="chartTitle">Average score vs log parameters</h2>
|
| 321 |
<p class="chart-sub" id="chartSub">Beautiful silver regression line over all featured models.</p>
|
|
|
|
|
|
|
|
|
|
|
|
|
| 322 |
</div>
|
| 323 |
<div class="badge" id="countBadge">0 models</div>
|
| 324 |
</div>
|
| 325 |
<canvas id="scatterChart"></canvas>
|
| 326 |
</section>
|
| 327 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 328 |
<section class="list-card">
|
| 329 |
<h2>Featured models</h2>
|
| 330 |
-
<div class="
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 331 |
</section>
|
| 332 |
</main>
|
| 333 |
</div>
|
| 334 |
</div>
|
| 335 |
|
| 336 |
<script>
|
| 337 |
-
// Edit this array privately whenever you want to add or remove models.
|
| 338 |
const MODELS = [
|
| 339 |
-
{ name: 'SmolLM2-135M', org: 'HuggingFace', params: 135000000,
|
| 340 |
-
{ name: 'GPT-X2-125M', org: 'AxiomicLabs', params: 125081664,
|
| 341 |
-
{ name: 'GPT-X-125M', org: 'AxiomicLabs', params: 124561728,
|
| 342 |
-
{ name: 'MobileLLM-R1-140M', org: 'Facebook', params: 140000000,
|
| 343 |
-
{ name: 'Supra-50M-Base', org: 'SupraLabs', params: 51786240,
|
| 344 |
-
{ name: '
|
| 345 |
-
|
| 346 |
-
{ name: '
|
| 347 |
-
{ name: '
|
| 348 |
-
{ name: '
|
| 349 |
-
{ name: '
|
| 350 |
-
{ name: 'GPT-
|
| 351 |
-
{ name: '
|
| 352 |
-
|
| 353 |
-
{ name: '
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 354 |
];
|
| 355 |
|
| 356 |
const BENCHMARKS = [
|
| 357 |
-
{ key: '
|
| 358 |
-
{ key: '
|
| 359 |
-
{ key: '
|
| 360 |
{ key: 'hellaswag', label: 'HellaSwag' },
|
| 361 |
-
{ key: 'piqa', label: 'PIQA' }
|
| 362 |
-
{ key: 'arithmark2', label: 'ArithMark-2' }
|
| 363 |
];
|
| 364 |
|
| 365 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 366 |
let chart = null;
|
| 367 |
|
| 368 |
-
const
|
| 369 |
if (n >= 1e9) return `${(n / 1e9).toFixed(2)}B`;
|
| 370 |
if (n >= 1e6) return `${(n / 1e6).toFixed(2)}M`;
|
| 371 |
if (n >= 1e3) return `${(n / 1e3).toFixed(2)}K`;
|
| 372 |
return String(n);
|
| 373 |
};
|
| 374 |
|
| 375 |
-
const
|
| 376 |
|
| 377 |
-
const
|
| 378 |
-
|
| 379 |
-
function linearRegression(points) {
|
| 380 |
const n = points.length;
|
| 381 |
let sumX = 0, sumY = 0, sumXY = 0, sumXX = 0;
|
| 382 |
for (const p of points) {
|
|
@@ -389,111 +655,143 @@
|
|
| 389 |
const slope = Math.abs(denom) < 1e-12 ? 0 : (n * sumXY - sumX * sumY) / denom;
|
| 390 |
const intercept = n ? (sumY - slope * sumX) / n : 0;
|
| 391 |
return { slope, intercept };
|
| 392 |
-
}
|
| 393 |
|
| 394 |
-
|
| 395 |
-
|
| 396 |
-
|
| 397 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 398 |
});
|
| 399 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 400 |
const rmse = Math.sqrt(mse);
|
| 401 |
return { mse, rmse, residuals };
|
| 402 |
-
}
|
| 403 |
-
|
| 404 |
-
function getChartData() {
|
| 405 |
-
return MODELS
|
| 406 |
-
.filter(m => Number.isFinite(m.params) && getScore(m, activeBenchmark) !== null)
|
| 407 |
-
.map(m => ({
|
| 408 |
-
name: m.name,
|
| 409 |
-
org: m.org,
|
| 410 |
-
params: m.params,
|
| 411 |
-
paramsDisplay: formatParams(m.params),
|
| 412 |
-
url: m.url,
|
| 413 |
-
x: Math.log10(m.params),
|
| 414 |
-
y: getScore(m, activeBenchmark)
|
| 415 |
-
}))
|
| 416 |
-
.sort((a, b) => a.x - b.x);
|
| 417 |
-
}
|
| 418 |
|
| 419 |
-
|
| 420 |
const box = document.getElementById('benchmarkButtons');
|
| 421 |
-
box.innerHTML = BENCHMARKS.map(b =>
|
| 422 |
-
<button class="${b.key === activeBenchmark ? 'active' : ''}" data-key="${b.key}">${b.label}</button>
|
| 423 |
-
|
| 424 |
box.querySelectorAll('button').forEach(btn => {
|
| 425 |
btn.addEventListener('click', () => {
|
| 426 |
activeBenchmark = btn.dataset.key;
|
| 427 |
render();
|
| 428 |
});
|
| 429 |
});
|
| 430 |
-
}
|
| 431 |
|
| 432 |
-
|
| 433 |
-
const
|
| 434 |
-
|
| 435 |
-
|
| 436 |
-
|
| 437 |
-
|
| 438 |
-
|
| 439 |
-
|
| 440 |
-
|
| 441 |
-
|
| 442 |
-
|
| 443 |
-
`).join('');
|
| 444 |
|
| 445 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 446 |
row.addEventListener('click', () => {
|
| 447 |
const url = row.dataset.url;
|
| 448 |
if (url) window.open(url, '_blank', 'noopener,noreferrer');
|
| 449 |
});
|
| 450 |
});
|
| 451 |
-
}
|
| 452 |
|
| 453 |
-
|
| 454 |
document.getElementById('statSlope').textContent = slope.toFixed(4);
|
| 455 |
document.getElementById('statIntercept').textContent = intercept.toFixed(2);
|
| 456 |
-
document.getElementById('
|
| 457 |
-
document.getElementById('
|
| 458 |
document.getElementById('countBadge').textContent = `${count} models`;
|
| 459 |
-
|
|
|
|
|
|
|
|
|
|
| 460 |
|
| 461 |
function render() {
|
| 462 |
renderBenchmarkButtons();
|
| 463 |
|
| 464 |
-
const data =
|
| 465 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 466 |
|
| 467 |
-
const points = data.map(d => ({ x: d.x, y: d.
|
| 468 |
const { slope, intercept } = linearRegression(points);
|
| 469 |
-
const { mse, rmse, residuals } = computeLoss(data,
|
| 470 |
-
|
|
|
|
|
|
|
|
|
|
| 471 |
|
| 472 |
updateStats(slope, intercept, mse, rmse, data.length);
|
| 473 |
-
renderModelList(data);
|
| 474 |
|
| 475 |
const xMin = Math.min(...data.map(d => d.x));
|
| 476 |
const xMax = Math.max(...data.map(d => d.x));
|
| 477 |
-
const
|
| 478 |
-
const xLow = xMin -
|
| 479 |
-
const xHigh = xMax +
|
| 480 |
const yLow = slope * xLow + intercept;
|
| 481 |
const yHigh = slope * xHigh + intercept;
|
| 482 |
-
|
| 483 |
-
const
|
|
|
|
| 484 |
|
| 485 |
const titleMap = {
|
| 486 |
-
|
| 487 |
-
|
| 488 |
-
|
| 489 |
hellaswag: 'HellaSwag vs log parameters',
|
| 490 |
-
piqa: 'PIQA vs log parameters'
|
| 491 |
-
arithmark2: 'ArithMark-2 vs log parameters'
|
| 492 |
};
|
| 493 |
|
| 494 |
-
document.getElementById('chartTitle').textContent = titleMap[activeBenchmark] ||
|
| 495 |
-
document.getElementById('chartSub').textContent = `
|
| 496 |
-
document.getElementById('fitNote').textContent = `Regression loss: MSE ${mse.toFixed(4)} · RMSE ${rmse.toFixed(2)}.`;
|
|
|
|
|
|
|
| 497 |
|
| 498 |
const ctx = document.getElementById('scatterChart').getContext('2d');
|
| 499 |
if (chart) chart.destroy();
|
|
@@ -510,7 +808,9 @@
|
|
| 510 |
{ x: xHigh, y: yHigh }
|
| 511 |
],
|
| 512 |
borderColor: 'rgba(236, 236, 241, 0.98)',
|
| 513 |
-
borderWidth: 3
|
|
|
|
|
|
|
| 514 |
pointRadius: 0,
|
| 515 |
tension: 0,
|
| 516 |
fill: false,
|
|
@@ -519,12 +819,26 @@
|
|
| 519 |
{
|
| 520 |
label: 'Models',
|
| 521 |
data: data.map(d => ({ ...d })),
|
|
|
|
| 522 |
pointRadius: 7,
|
| 523 |
pointHoverRadius: 10,
|
|
|
|
| 524 |
borderWidth: 1.5,
|
| 525 |
-
backgroundColor:
|
| 526 |
-
|
| 527 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 528 |
hoverBorderColor: 'rgba(255,255,255,0.98)',
|
| 529 |
order: 2
|
| 530 |
}
|
|
@@ -535,6 +849,9 @@
|
|
| 535 |
maintainAspectRatio: false,
|
| 536 |
animation: { duration: 0 },
|
| 537 |
interaction: { mode: 'nearest', intersect: true },
|
|
|
|
|
|
|
|
|
|
| 538 |
onClick: (event, elements) => {
|
| 539 |
if (!elements.length) return;
|
| 540 |
const hit = elements[0];
|
|
@@ -545,11 +862,11 @@
|
|
| 545 |
plugins: {
|
| 546 |
legend: { display: false },
|
| 547 |
tooltip: {
|
| 548 |
-
backgroundColor: 'rgba(
|
| 549 |
-
borderColor: 'rgba(236,236,241,0.
|
| 550 |
borderWidth: 1,
|
| 551 |
titleColor: '#ffffff',
|
| 552 |
-
bodyColor: '#
|
| 553 |
padding: 12,
|
| 554 |
displayColors: false,
|
| 555 |
callbacks: {
|
|
@@ -559,8 +876,8 @@
|
|
| 559 |
const pred = slope * d.x + intercept;
|
| 560 |
const resid = d.y - pred;
|
| 561 |
return [
|
| 562 |
-
`Params: ${
|
| 563 |
-
`Score: ${d.
|
| 564 |
`Residual: ${resid >= 0 ? '+' : ''}${resid.toFixed(2)} pts`,
|
| 565 |
`Prediction: ${pred.toFixed(2)}%`
|
| 566 |
];
|
|
@@ -575,13 +892,13 @@
|
|
| 575 |
max: xHigh,
|
| 576 |
grid: { color: 'rgba(255,255,255,0.045)' },
|
| 577 |
ticks: {
|
| 578 |
-
color: '#
|
| 579 |
callback: v => `10^${Number(v).toFixed(1)}`
|
| 580 |
},
|
| 581 |
title: {
|
| 582 |
display: true,
|
| 583 |
text: 'Log₁₀(parameters)',
|
| 584 |
-
color: '#
|
| 585 |
font: { weight: '600' }
|
| 586 |
}
|
| 587 |
},
|
|
@@ -590,13 +907,13 @@
|
|
| 590 |
max: yMax,
|
| 591 |
grid: { color: 'rgba(255,255,255,0.045)' },
|
| 592 |
ticks: {
|
| 593 |
-
color: '#
|
| 594 |
callback: v => `${Number(v).toFixed(0)}%`
|
| 595 |
},
|
| 596 |
title: {
|
| 597 |
display: true,
|
| 598 |
text: 'Score (%)',
|
| 599 |
-
color: '#
|
| 600 |
font: { weight: '600' }
|
| 601 |
}
|
| 602 |
}
|
|
@@ -605,7 +922,19 @@
|
|
| 605 |
});
|
| 606 |
}
|
| 607 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 608 |
render();
|
| 609 |
</script>
|
| 610 |
</body>
|
| 611 |
-
</html>
|
|
|
|
| 7 |
<script src="https://cdn.jsdelivr.net/npm/chart.js@4"></script>
|
| 8 |
<style>
|
| 9 |
:root {
|
| 10 |
+
--bg: #08080a;
|
| 11 |
+
--bg2: #101014;
|
| 12 |
+
--surface: rgba(255,255,255,0.035);
|
| 13 |
+
--surface-2: rgba(255,255,255,0.05);
|
| 14 |
--border: rgba(255,255,255,0.08);
|
| 15 |
+
--border-2: rgba(255,255,255,0.13);
|
| 16 |
+
--text: #f3f3f5;
|
| 17 |
+
--muted: #a0a0aa;
|
| 18 |
+
--muted-2: #73737d;
|
| 19 |
+
--silver: #e0e2e8;
|
| 20 |
+
--silver-2: #c2c6cf;
|
| 21 |
+
--silver-3: #8f949f;
|
| 22 |
+
--shadow: 0 24px 70px rgba(0,0,0,0.48);
|
| 23 |
+
--radius: 24px;
|
| 24 |
}
|
| 25 |
|
| 26 |
* { box-sizing: border-box; }
|
|
|
|
| 32 |
font-family: Inter, ui-sans-serif, system-ui, -apple-system, BlinkMacSystemFont, "Segoe UI", sans-serif;
|
| 33 |
color: var(--text);
|
| 34 |
background:
|
| 35 |
+
radial-gradient(circle at top left, rgba(255,255,255,0.08), transparent 27%),
|
| 36 |
radial-gradient(circle at bottom right, rgba(255,255,255,0.05), transparent 24%),
|
| 37 |
+
linear-gradient(180deg, #08080a 0%, #060607 100%);
|
| 38 |
}
|
| 39 |
|
| 40 |
.app {
|
| 41 |
+
max-width: 1440px;
|
| 42 |
margin: 0 auto;
|
| 43 |
display: grid;
|
| 44 |
gap: 18px;
|
| 45 |
}
|
| 46 |
|
| 47 |
+
.hero,
|
| 48 |
+
.panel,
|
| 49 |
+
.chart-card,
|
| 50 |
+
.list-card,
|
| 51 |
+
.info-card {
|
| 52 |
+
background: linear-gradient(180deg, rgba(255,255,255,0.04), rgba(255,255,255,0.02));
|
| 53 |
border: 1px solid var(--border);
|
| 54 |
box-shadow: var(--shadow);
|
| 55 |
+
backdrop-filter: blur(12px);
|
| 56 |
}
|
| 57 |
|
| 58 |
.hero {
|
| 59 |
+
border-radius: calc(var(--radius) + 6px);
|
| 60 |
+
padding: 26px;
|
| 61 |
position: relative;
|
| 62 |
overflow: hidden;
|
| 63 |
}
|
|
|
|
| 66 |
content: "";
|
| 67 |
position: absolute;
|
| 68 |
inset: 0;
|
| 69 |
+
background: linear-gradient(90deg, rgba(255,255,255,0.03), transparent 38%, rgba(255,255,255,0.02));
|
| 70 |
pointer-events: none;
|
| 71 |
}
|
| 72 |
|
| 73 |
.eyebrow {
|
| 74 |
+
color: var(--silver-2);
|
| 75 |
text-transform: uppercase;
|
| 76 |
letter-spacing: 0.22em;
|
| 77 |
font-size: 11px;
|
|
|
|
| 80 |
|
| 81 |
h1 {
|
| 82 |
margin: 0;
|
| 83 |
+
font-size: clamp(34px, 5vw, 68px);
|
| 84 |
line-height: 0.96;
|
| 85 |
letter-spacing: -0.06em;
|
| 86 |
font-weight: 850;
|
| 87 |
}
|
| 88 |
|
| 89 |
+
.title-silver {
|
| 90 |
+
color: var(--silver);
|
| 91 |
+
text-shadow: 0 0 22px rgba(255,255,255,0.10);
|
| 92 |
+
}
|
| 93 |
+
|
| 94 |
.sub {
|
| 95 |
margin: 14px 0 0;
|
| 96 |
color: var(--muted);
|
| 97 |
+
max-width: 980px;
|
| 98 |
line-height: 1.7;
|
| 99 |
font-size: 14px;
|
| 100 |
}
|
| 101 |
|
| 102 |
.layout {
|
| 103 |
display: grid;
|
| 104 |
+
grid-template-columns: 330px minmax(0, 1fr);
|
| 105 |
gap: 18px;
|
| 106 |
align-items: start;
|
| 107 |
}
|
|
|
|
| 113 |
top: 18px;
|
| 114 |
}
|
| 115 |
|
| 116 |
+
.panel h2,
|
| 117 |
+
.chart-card h2,
|
| 118 |
+
.list-card h2,
|
| 119 |
+
.info-card h2 {
|
| 120 |
margin: 0 0 14px;
|
| 121 |
font-size: 14px;
|
| 122 |
letter-spacing: 0.12em;
|
|
|
|
| 130 |
.section-label {
|
| 131 |
display: block;
|
| 132 |
margin-bottom: 10px;
|
| 133 |
+
color: var(--silver-2);
|
| 134 |
font-size: 12px;
|
| 135 |
text-transform: uppercase;
|
| 136 |
letter-spacing: 0.12em;
|
|
|
|
| 142 |
gap: 8px;
|
| 143 |
}
|
| 144 |
|
| 145 |
+
button,
|
| 146 |
+
.chip {
|
| 147 |
border: 1px solid rgba(255,255,255,0.10);
|
| 148 |
background: rgba(255,255,255,0.03);
|
| 149 |
color: var(--muted);
|
|
|
|
| 153 |
transition: all 0.15s ease;
|
| 154 |
font: inherit;
|
| 155 |
font-size: 13px;
|
| 156 |
+
user-select: none;
|
| 157 |
}
|
| 158 |
|
| 159 |
+
button:hover,
|
| 160 |
+
.chip:hover {
|
| 161 |
border-color: rgba(255,255,255,0.24);
|
| 162 |
color: var(--text);
|
| 163 |
transform: translateY(-1px);
|
| 164 |
}
|
| 165 |
|
| 166 |
+
button.active,
|
| 167 |
+
.chip.active {
|
| 168 |
color: #fafafa;
|
| 169 |
+
border-color: rgba(224,226,232,0.48);
|
| 170 |
+
background: linear-gradient(180deg, rgba(224,226,232,0.18), rgba(224,226,232,0.08));
|
| 171 |
box-shadow: 0 0 0 1px rgba(255,255,255,0.03) inset;
|
| 172 |
}
|
| 173 |
|
|
|
|
| 186 |
}
|
| 187 |
|
| 188 |
.stat .k {
|
| 189 |
+
color: var(--muted-2);
|
| 190 |
font-size: 11px;
|
| 191 |
text-transform: uppercase;
|
| 192 |
letter-spacing: 0.12em;
|
|
|
|
| 194 |
|
| 195 |
.stat .v {
|
| 196 |
margin-top: 6px;
|
| 197 |
+
font-size: 19px;
|
| 198 |
+
font-weight: 760;
|
| 199 |
color: var(--silver);
|
| 200 |
font-variant-numeric: tabular-nums;
|
| 201 |
}
|
| 202 |
|
| 203 |
.small-note {
|
| 204 |
+
color: var(--muted-2);
|
| 205 |
font-size: 12px;
|
| 206 |
line-height: 1.6;
|
| 207 |
margin-top: 10px;
|
|
|
|
| 209 |
|
| 210 |
.chart-card {
|
| 211 |
border-radius: var(--radius);
|
| 212 |
+
padding: 18px 18px 12px;
|
| 213 |
+
min-height: 690px;
|
| 214 |
+
position: relative;
|
| 215 |
+
overflow: hidden;
|
| 216 |
}
|
| 217 |
|
| 218 |
.chart-top {
|
| 219 |
display: flex;
|
| 220 |
+
align-items: flex-start;
|
| 221 |
justify-content: space-between;
|
| 222 |
gap: 12px;
|
| 223 |
margin-bottom: 12px;
|
|
|
|
| 226 |
|
| 227 |
.chart-title {
|
| 228 |
margin: 0;
|
| 229 |
+
font-size: 19px;
|
| 230 |
+
font-weight: 760;
|
| 231 |
letter-spacing: -0.03em;
|
| 232 |
}
|
| 233 |
|
|
|
|
| 235 |
margin: 4px 0 0;
|
| 236 |
color: var(--muted);
|
| 237 |
font-size: 13px;
|
| 238 |
+
line-height: 1.5;
|
| 239 |
}
|
| 240 |
|
| 241 |
.badge {
|
| 242 |
+
padding: 9px 12px;
|
| 243 |
border-radius: 999px;
|
| 244 |
+
background: rgba(224,226,232,0.08);
|
| 245 |
+
border: 1px solid rgba(224,226,232,0.16);
|
| 246 |
color: var(--silver);
|
| 247 |
font-size: 13px;
|
| 248 |
white-space: nowrap;
|
| 249 |
+
font-variant-numeric: tabular-nums;
|
| 250 |
}
|
| 251 |
|
| 252 |
+
.callout {
|
| 253 |
+
margin-top: 10px;
|
| 254 |
+
display: inline-flex;
|
| 255 |
+
align-items: center;
|
| 256 |
+
gap: 10px;
|
| 257 |
+
padding: 9px 12px;
|
| 258 |
+
border-radius: 999px;
|
| 259 |
+
border: 1px solid rgba(224,226,232,0.16);
|
| 260 |
+
background: rgba(224,226,232,0.06);
|
| 261 |
+
color: #efeff2;
|
| 262 |
+
font-size: 12px;
|
| 263 |
+
letter-spacing: 0.01em;
|
| 264 |
+
}
|
| 265 |
+
|
| 266 |
+
.callout .dot {
|
| 267 |
+
width: 9px;
|
| 268 |
+
height: 9px;
|
| 269 |
+
border-radius: 50%;
|
| 270 |
+
background: linear-gradient(180deg, #f7f7f9, #b9bec8);
|
| 271 |
+
box-shadow: 0 0 14px rgba(255,255,255,0.22);
|
| 272 |
+
flex: 0 0 auto;
|
| 273 |
+
}
|
| 274 |
+
|
| 275 |
+
.filters {
|
| 276 |
+
display: flex;
|
| 277 |
+
flex-wrap: wrap;
|
| 278 |
+
gap: 8px;
|
| 279 |
+
margin: 10px 0 16px;
|
| 280 |
+
}
|
| 281 |
+
|
| 282 |
+
.search {
|
| 283 |
+
width: min(100%, 420px);
|
| 284 |
+
border: 1px solid rgba(255,255,255,0.10);
|
| 285 |
+
background: rgba(255,255,255,0.03);
|
| 286 |
+
color: var(--text);
|
| 287 |
+
border-radius: 16px;
|
| 288 |
+
padding: 12px 14px;
|
| 289 |
+
outline: none;
|
| 290 |
+
font: inherit;
|
| 291 |
+
transition: border-color 0.15s ease, background 0.15s ease;
|
| 292 |
+
}
|
| 293 |
+
|
| 294 |
+
.search::placeholder { color: var(--muted-2); }
|
| 295 |
+
|
| 296 |
+
.search:focus {
|
| 297 |
+
border-color: rgba(224,226,232,0.28);
|
| 298 |
+
background: rgba(255,255,255,0.05);
|
| 299 |
+
}
|
| 300 |
+
|
| 301 |
+
canvas {
|
| 302 |
+
width: 100% !important;
|
| 303 |
+
height: 560px !important;
|
| 304 |
+
}
|
| 305 |
+
|
| 306 |
+
.info-grid {
|
| 307 |
+
display: grid;
|
| 308 |
+
grid-template-columns: repeat(3, minmax(0, 1fr));
|
| 309 |
+
gap: 14px;
|
| 310 |
+
}
|
| 311 |
+
|
| 312 |
+
.info-card {
|
| 313 |
+
border-radius: 20px;
|
| 314 |
+
padding: 16px;
|
| 315 |
+
}
|
| 316 |
+
|
| 317 |
+
.info-card .big {
|
| 318 |
+
font-size: 22px;
|
| 319 |
+
font-weight: 800;
|
| 320 |
+
color: var(--silver);
|
| 321 |
+
margin-top: 6px;
|
| 322 |
+
font-variant-numeric: tabular-nums;
|
| 323 |
+
}
|
| 324 |
+
|
| 325 |
+
.info-card .label {
|
| 326 |
+
color: var(--muted-2);
|
| 327 |
+
font-size: 11px;
|
| 328 |
+
text-transform: uppercase;
|
| 329 |
+
letter-spacing: 0.12em;
|
| 330 |
+
}
|
| 331 |
|
| 332 |
.list-card {
|
| 333 |
border-radius: var(--radius);
|
| 334 |
padding: 16px;
|
| 335 |
}
|
| 336 |
|
| 337 |
+
.table-wrap {
|
|
|
|
|
|
|
|
|
|
| 338 |
overflow: auto;
|
| 339 |
+
border-radius: 18px;
|
| 340 |
+
border: 1px solid rgba(255,255,255,0.08);
|
| 341 |
+
background: rgba(255,255,255,0.02);
|
| 342 |
}
|
| 343 |
|
| 344 |
+
table {
|
| 345 |
+
width: 100%;
|
| 346 |
+
border-collapse: collapse;
|
| 347 |
+
min-width: 900px;
|
| 348 |
+
}
|
| 349 |
+
|
| 350 |
+
thead th {
|
| 351 |
+
position: sticky;
|
| 352 |
+
top: 0;
|
| 353 |
+
background: rgba(12,12,14,0.96);
|
| 354 |
+
backdrop-filter: blur(8px);
|
| 355 |
+
color: var(--muted);
|
| 356 |
+
text-align: left;
|
| 357 |
+
font-size: 10px;
|
| 358 |
+
text-transform: uppercase;
|
| 359 |
+
letter-spacing: 0.12em;
|
| 360 |
+
padding: 14px 14px;
|
| 361 |
+
border-bottom: 1px solid rgba(255,255,255,0.08);
|
| 362 |
+
cursor: pointer;
|
| 363 |
+
user-select: none;
|
| 364 |
+
white-space: nowrap;
|
| 365 |
+
}
|
| 366 |
+
|
| 367 |
+
thead th:hover { color: var(--silver); }
|
| 368 |
+
|
| 369 |
+
tbody tr {
|
| 370 |
+
border-bottom: 1px solid rgba(255,255,255,0.06);
|
| 371 |
+
transition: background 0.12s ease, transform 0.12s ease;
|
| 372 |
cursor: pointer;
|
| 373 |
}
|
| 374 |
|
| 375 |
+
tbody tr:hover {
|
| 376 |
+
background: rgba(224,226,232,0.05);
|
|
|
|
|
|
|
| 377 |
}
|
| 378 |
|
| 379 |
+
tbody td {
|
| 380 |
+
padding: 13px 14px;
|
| 381 |
+
color: var(--text);
|
| 382 |
+
font-size: 13px;
|
| 383 |
+
vertical-align: middle;
|
| 384 |
+
}
|
| 385 |
|
| 386 |
+
tbody td.muted {
|
| 387 |
+
color: var(--muted);
|
|
|
|
|
|
|
| 388 |
}
|
| 389 |
|
| 390 |
+
.above {
|
| 391 |
+
color: #f2f2f6;
|
| 392 |
+
text-shadow: 0 0 10px rgba(255,255,255,0.08);
|
| 393 |
+
}
|
| 394 |
+
|
| 395 |
+
.below {
|
| 396 |
+
color: #8f949d;
|
| 397 |
+
}
|
| 398 |
+
|
| 399 |
+
.pill-up {
|
| 400 |
+
display: inline-flex;
|
| 401 |
+
align-items: center;
|
| 402 |
+
gap: 6px;
|
| 403 |
+
padding: 5px 10px;
|
| 404 |
+
border-radius: 999px;
|
| 405 |
+
border: 1px solid rgba(224,226,232,0.18);
|
| 406 |
+
background: rgba(224,226,232,0.08);
|
| 407 |
+
color: var(--silver);
|
| 408 |
+
font-size: 11px;
|
| 409 |
+
font-weight: 700;
|
| 410 |
+
letter-spacing: 0.03em;
|
| 411 |
+
}
|
| 412 |
+
|
| 413 |
+
.pill-down {
|
| 414 |
+
display: inline-flex;
|
| 415 |
+
align-items: center;
|
| 416 |
+
gap: 6px;
|
| 417 |
+
padding: 5px 10px;
|
| 418 |
+
border-radius: 999px;
|
| 419 |
+
border: 1px solid rgba(255,255,255,0.08);
|
| 420 |
+
background: rgba(255,255,255,0.03);
|
| 421 |
+
color: var(--muted);
|
| 422 |
+
font-size: 11px;
|
| 423 |
+
font-weight: 700;
|
| 424 |
+
letter-spacing: 0.03em;
|
| 425 |
+
}
|
| 426 |
+
|
| 427 |
+
.footer-note {
|
| 428 |
+
color: var(--muted-2);
|
| 429 |
+
font-size: 12px;
|
| 430 |
+
line-height: 1.6;
|
| 431 |
+
padding-top: 2px;
|
| 432 |
+
}
|
| 433 |
+
|
| 434 |
+
@media (max-width: 1120px) {
|
| 435 |
+
.layout {
|
| 436 |
+
grid-template-columns: 1fr;
|
| 437 |
+
}
|
| 438 |
+
.panel {
|
| 439 |
+
position: static;
|
| 440 |
+
}
|
| 441 |
+
canvas { height: 520px !important; }
|
| 442 |
+
.info-grid { grid-template-columns: 1fr; }
|
| 443 |
+
}
|
| 444 |
+
|
| 445 |
+
@media (max-width: 760px) {
|
| 446 |
body { padding: 16px; }
|
| 447 |
+
.hero, .panel, .chart-card, .list-card, .info-card { border-radius: 18px; }
|
| 448 |
.stat-grid { grid-template-columns: 1fr; }
|
|
|
|
| 449 |
canvas { height: 420px !important; }
|
| 450 |
+
table { min-width: 760px; }
|
| 451 |
}
|
| 452 |
</style>
|
| 453 |
</head>
|
|
|
|
| 457 |
<div class="eyebrow">Regression explorer</div>
|
| 458 |
<h1><span class="title-silver">Silver</span> benchmark fit line</h1>
|
| 459 |
<p class="sub">
|
| 460 |
+
Pick a benchmark, fit a linear regression on <b>log10(parameters)</b>, and inspect how every model sits against the size trend.
|
| 461 |
+
Hover any dot to see the model name. Click a dot to open the model page.
|
| 462 |
</p>
|
| 463 |
</section>
|
| 464 |
|
|
|
|
| 471 |
<div class="btn-grid" id="benchmarkButtons"></div>
|
| 472 |
</div>
|
| 473 |
|
| 474 |
+
<div class="section">
|
| 475 |
+
<span class="section-label">Search models</span>
|
| 476 |
+
<input id="searchBox" class="search" type="text" placeholder="Type a model or org name..." />
|
| 477 |
+
<div class="small-note">Filters the chart and the table at once.</div>
|
| 478 |
+
</div>
|
| 479 |
+
|
| 480 |
<div class="section">
|
| 481 |
<span class="section-label">Regression stats</span>
|
| 482 |
<div class="stat-grid">
|
| 483 |
<div class="stat"><div class="k">Slope</div><div class="v" id="statSlope">—</div></div>
|
| 484 |
<div class="stat"><div class="k">Intercept</div><div class="v" id="statIntercept">—</div></div>
|
| 485 |
+
<div class="stat"><div class="k">MSE</div><div class="v" id="statMSE">—</div></div>
|
| 486 |
+
<div class="stat"><div class="k">RMSE</div><div class="v" id="statRMSE">—</div></div>
|
| 487 |
</div>
|
| 488 |
<div class="small-note" id="fitNote">The fit updates instantly when you switch benchmarks.</div>
|
| 489 |
</div>
|
| 490 |
+
|
| 491 |
+
<div class="section">
|
| 492 |
+
<span class="section-label">Legend</span>
|
| 493 |
+
<div class="small-note">
|
| 494 |
+
Silver points above the line are outperforming the size trend. Gray points are below it.
|
| 495 |
+
</div>
|
| 496 |
+
</div>
|
| 497 |
</aside>
|
| 498 |
|
| 499 |
<main style="display:grid; gap:18px;">
|
|
|
|
| 502 |
<div>
|
| 503 |
<h2 class="chart-title" id="chartTitle">Average score vs log parameters</h2>
|
| 504 |
<p class="chart-sub" id="chartSub">Beautiful silver regression line over all featured models.</p>
|
| 505 |
+
<div class="callout">
|
| 506 |
+
<span class="dot"></span>
|
| 507 |
+
<span>Above the dotted line = above expected for that parameter scale.</span>
|
| 508 |
+
</div>
|
| 509 |
</div>
|
| 510 |
<div class="badge" id="countBadge">0 models</div>
|
| 511 |
</div>
|
| 512 |
<canvas id="scatterChart"></canvas>
|
| 513 |
</section>
|
| 514 |
|
| 515 |
+
<section class="info-grid">
|
| 516 |
+
<div class="info-card">
|
| 517 |
+
<div class="label">Selected benchmark</div>
|
| 518 |
+
<div class="big" id="infoBenchmark">Avg</div>
|
| 519 |
+
</div>
|
| 520 |
+
<div class="info-card">
|
| 521 |
+
<div class="label">Visible models</div>
|
| 522 |
+
<div class="big" id="infoCount">0</div>
|
| 523 |
+
</div>
|
| 524 |
+
<div class="info-card">
|
| 525 |
+
<div class="label">Regression loss</div>
|
| 526 |
+
<div class="big" id="infoLoss">—</div>
|
| 527 |
+
</div>
|
| 528 |
+
</section>
|
| 529 |
+
|
| 530 |
<section class="list-card">
|
| 531 |
<h2>Featured models</h2>
|
| 532 |
+
<div class="table-wrap">
|
| 533 |
+
<table>
|
| 534 |
+
<thead>
|
| 535 |
+
<tr>
|
| 536 |
+
<th data-sort="name">Model</th>
|
| 537 |
+
<th data-sort="org">Org</th>
|
| 538 |
+
<th data-sort="params">Params</th>
|
| 539 |
+
<th data-sort="score">Score</th>
|
| 540 |
+
<th data-sort="residual">Residual</th>
|
| 541 |
+
<th data-sort="side">Relation</th>
|
| 542 |
+
</tr>
|
| 543 |
+
</thead>
|
| 544 |
+
<tbody id="modelTable"></tbody>
|
| 545 |
+
</table>
|
| 546 |
+
</div>
|
| 547 |
+
<div class="footer-note">
|
| 548 |
+
Clicking a row opens the model page.
|
| 549 |
+
</div>
|
| 550 |
</section>
|
| 551 |
</main>
|
| 552 |
</div>
|
| 553 |
</div>
|
| 554 |
|
| 555 |
<script>
|
|
|
|
| 556 |
const MODELS = [
|
| 557 |
+
{ name: 'SmolLM2-135M', org: 'HuggingFace', params: 135000000, arc_easy: 0.5858585858585859, arc_challenge: 0.2960750853242321, piqa: 0.6849836779107725, hellaswag: 0.43108942441744674, avg_4: 0.5177858993388165, url: 'https://huggingface.co/HuggingFaceTB/SmolLM2-135M' },
|
| 558 |
+
{ name: 'GPT-X2-125M', org: 'AxiomicLabs', params: 125081664, arc_easy: 0.5151515151515151, arc_challenge: 0.2764505119453925, piqa: 0.6751904243743199, hellaswag: 0.40529774945230035, avg_4: 0.4898234185711896, url: 'https://huggingface.co/AxiomicLabs/GPT-X2-125M' },
|
| 559 |
+
{ name: 'GPT-X-125M', org: 'AxiomicLabs', params: 124561728, arc_easy: 0.5075757575757576, arc_challenge: 0.2687713310580205, piqa: 0.6517954298150164, hellaswag: 0.36496713802031466, avg_4: 0.464060299783275, url: 'https://huggingface.co/AxiomicLabs/GPT-X-125M' },
|
| 560 |
+
{ name: 'MobileLLM-R1-140M-base', org: 'Facebook', params: 140000000, arc_easy: 0.4992424242424242, arc_challenge: 0.24744027303754265, piqa: 0.6322089227421109, hellaswag: 0.3384379853648824, avg_4: 0.443830, url: 'https://huggingface.co/facebook/MobileLLM-R1-140M-base' },
|
| 561 |
+
{ name: 'Supra-50M-Base', org: 'SupraLabs', params: 51786240, arc_easy: 0.460016835016835, arc_challenge: 0.25, piqa: 0.6207834602829162, hellaswag: 0.31776538538139815, avg_4: 0.4252144857313513, url: 'https://huggingface.co/SupraLabs/Supra-50M-Base' },
|
| 562 |
+
{ name: 'Shards-1', org: 'CompactAI', params: 54500000, arc_easy: 0.4112, arc_challenge: 0.2099, piqa: 0.5822, hellaswag: 0.2920, avg_4: 0.3698, url: 'https://huggingface.co/CompactAI-O/Shard-1' },
|
| 563 |
+
|
| 564 |
+
{ name: 'Supra-50M-Instruct', org: 'SupraLabs', params: 51786240, arc_easy: 0.4440, arc_challenge: 0.2730, piqa: 0.5947, hellaswag: 0.2909, avg_4: 0.4030, url: 'https://huggingface.co/SupraLabs/Supra-50M-Instruct' },
|
| 565 |
+
{ name: 'Supra-50M-Reasoning', org: 'SupraLabs', params: 51786240, arc_easy: 0.4444, arc_challenge: 0.2739, piqa: 0.5930, hellaswag: 0.2910, avg_4: 0.4006, url: 'https://huggingface.co/SupraLabs/Supra-50M-Reasoning' },
|
| 566 |
+
{ name: 'SmolLM-135M', org: 'HuggingFace', params: 135000000, arc_easy: 0.5631, arc_challenge: 0.2901, piqa: 0.6828, hellaswag: 0.4270, avg_4: 0.4908, url: 'https://huggingface.co/HuggingFaceTB/SmolLM-135M' },
|
| 567 |
+
{ name: 'OPT-125M', org: 'Facebook', params: 125000000, arc_easy: 0.4028, arc_challenge: 0.2270, piqa: 0.6224, hellaswag: 0.3131, avg_4: 0.3913, url: 'https://huggingface.co/facebook/opt-125m' },
|
| 568 |
+
{ name: 'GPT-2', org: 'OpenAI', params: 124000000, arc_easy: 0.3935, arc_challenge: 0.2235, piqa: 0.6208, hellaswag: 0.3126, avg_4: 0.3892, url: 'https://huggingface.co/openai-community/gpt2' },
|
| 569 |
+
{ name: 'Spark-5M-Base-v4', org: 'LH-Tech-AI', params: 5000000, arc_easy: 0.3316, arc_challenge: 0.2150, piqa: 0.5332, hellaswag: 0.2703, avg_4: 0.3250, url: 'https://huggingface.co/LH-Tech-AI/Spark-5M-Base-v4' },
|
| 570 |
+
|
| 571 |
+
{ name: 'Supra-Mini-v5-8M', org: 'SupraLabs', params: 7870000, arc_easy: 0.3321, arc_challenge: 0.2116, piqa: 0.5403, hellaswag: 0.2637, avg_4: 0.3219, url: 'https://huggingface.co/SupraLabs/Supra-Mini-v5-8M' },
|
| 572 |
+
{ name: 'Pythia-70M', org: 'EleutherAI', params: 70000000, arc_easy: 0.3165, arc_challenge: 0.2363, piqa: 0.5348, hellaswag: 0.2749, avg_4: 0.3657, url: 'https://huggingface.co/EleutherAI/pythia-70m' },
|
| 573 |
+
{ name: 'Supra-Mini-v4-2M', org: 'SupraLabs', params: 2620000, arc_easy: 0.3098, arc_challenge: 0.2150, piqa: 0.5190, hellaswag: 0.2552, avg_4: 0.3188, url: 'https://huggingface.co/SupraLabs/Supra-Mini-v4-2M' },
|
| 574 |
+
{ name: 'Pythia-31M', org: 'EleutherAI', params: 31000000, arc_easy: 0.3388, arc_challenge: 0.2167, piqa: 0.5626, hellaswag: 0.2714, avg_4: 0.3474, url: 'https://huggingface.co/EleutherAI/pythia-31m' },
|
| 575 |
+
{ name: 'Stentor3-50M', org: 'StentorLabs', params: 50000000, arc_easy: 0.2967, arc_challenge: 0.2167, piqa: 0.5375, hellaswag: 0.2710, avg_4: 0.3335, url: 'https://huggingface.co/StentorLabs/Stentor3-50M' },
|
| 576 |
+
{ name: 'Stentor3-20M', org: 'StentorLabs', params: 20000000, arc_easy: 0.2950, arc_challenge: 0.2312, piqa: 0.5506, hellaswag: 0.2706, avg_4: 0.3394, url: 'https://huggingface.co/StentorLabs/Stentor3-20M' },
|
| 577 |
+
{ name: 'Portimbria-150M', org: 'StentorLabs', params: 151026432, arc_easy: 0.3582, arc_challenge: 0.1877, piqa: 0.5827, hellaswag: 0.2709, avg_4: 0.3499, url: 'https://huggingface.co/StentorLabs/Portimbria-150M' },
|
| 578 |
+
{ name: 'nanowhale-100m-base', org: 'HuggingFace', params: 100000000, arc_easy: 0.2879, arc_challenge: 0.2483, piqa: 0.5180, hellaswag: 0.2631, avg_4: 0.3293, url: 'https://huggingface.co/HuggingFaceTB/nanowhale-100m-base' },
|
| 579 |
+
{ name: 'Pythia-14M', org: 'EleutherAI', params: 14000000, arc_easy: 0.3228, arc_challenge: 0.2100, piqa: 0.5588, hellaswag: 0.2620, avg_4: 0.3384, url: 'https://huggingface.co/EleutherAI/pythia-14m' },
|
| 580 |
+
{ name: 'Tenete-8M', org: 'Harley-ml', params: 8000000, arc_easy: 0.3169, arc_challenge: 0.2184, piqa: 0.5566, hellaswag: 0.2675, avg_4: 0.3409, url: 'https://huggingface.co/Harley-ml/Tenete-8M' },
|
| 581 |
+
{ name: 'Dillion-1.2M', org: 'Harley-ml', params: 1281384, arc_easy: 0.3119, arc_challenge: 0.2278, piqa: 0.5305, hellaswag: 0.2665, avg_4: 0.3338, url: 'https://huggingface.co/Harley-ml/Dillion-1.2M' },
|
| 582 |
+
{ name: 'CinnabarLM-1.4M-Base', org: 'MihaiPopa-1', params: 1510000, arc_easy: 0.2854, arc_challenge: 0.2338, piqa: 0.5250, hellaswag: 0.2708, avg_4: 0.3416, url: 'https://huggingface.co/MihaiPopa-1/CinnabarLM-1.4M-Base' },
|
| 583 |
+
{ name: 'CinnabarLM-4M-Base', org: 'MihaiPopa-1', params: 4230000, arc_easy: 0.2828, arc_challenge: 0.2270, piqa: 0.5229, hellaswag: 0.2771, avg_4: 0.3302, url: 'https://huggingface.co/MihaiPopa-1/CinnabarLM-4M-Base' },
|
| 584 |
+
{ name: 'CinnabarLM-1.5M-Base', org: 'MihaiPopa-1', params: 1710000, arc_easy: 0.2811, arc_challenge: 0.2193, piqa: 0.5294, hellaswag: 0.2708, avg_4: 0.3338, url: 'https://huggingface.co/MihaiPopa-1/CinnabarLM-1.5M-Base' },
|
| 585 |
+
{ name: 'Dillionv2-1.3M', org: 'Harley-ml', params: 1285200, arc_easy: 0.2963, arc_challenge: 0.2224, piqa: 0.5305, hellaswag: 0.2738, avg_4: 0.3292, url: 'https://huggingface.co/Harley-ml/Dillionv2-1.3M' },
|
| 586 |
+
{ name: 'KeyLM-75M', org: 'Eclipse-Senpai', params: 75251200, arc_easy: 0.3573, arc_challenge: 0.2398, piqa: 0.6050, hellaswag: 0.2966, avg_4: 0.3947, url: 'https://huggingface.co/Eclipse-Senpai/KeyLM-75M' },
|
| 587 |
+
{ name: 'Supra-Mini-v6-1M', org: 'SupraLabs', params: 1410688, arc_easy: 0.3068, arc_challenge: 0.2048, piqa: 0.5370, hellaswag: 0.2722, avg_4: 0.3322, url: 'https://huggingface.co/SupraLabs/Supra-Mini-v6-1M' },
|
| 588 |
+
{ name: 'GPT-S-1.4M', org: 'AxiomicLabs', params: 1426000, arc_easy: 0.3186, arc_challenge: 0.2210, piqa: 0.5511, hellaswag: 0.2686, avg_4: 0.3341, url: 'https://huggingface.co/AxiomicLabs/GPT-S-1.4M' },
|
| 589 |
+
{ name: 'Archaea-74M', org: 'GODELEV', params: 74016256, arc_easy: 0.3906, arc_challenge: 0.2270, piqa: 0.5827, hellaswag: 0.2727, avg_4: 0.3812, url: 'https://huggingface.co/GODELEV/Archaea-74M' },
|
| 590 |
+
{ name: 'Cali-0.1B', org: 'Sandroeth', params: 123782400, arc_easy: 0.2753, arc_challenge: 0.2449, piqa: 0.5212, hellaswag: 0.2684, avg_4: 0.3274, url: 'https://huggingface.co/Sandroeth/cali-0.1B' },
|
| 591 |
+
{ name: 'Quark-50M', org: 'ThingAI', params: 56666496, arc_easy: 0.3678, arc_challenge: 0.2526, piqa: 0.5783, hellaswag: 0.2848, avg_4: 0.3717, url: 'https://huggingface.co/ThingAI/Quark-50m' },
|
| 592 |
+
{ name: 'Quark-135M', org: 'ThingAI', params: 134561088, arc_easy: 0.4789, arc_challenge: 0.2824, piqa: 0.5865, hellaswag: 0.3137, avg_4: 0.4143, url: 'https://huggingface.co/ThingAI/Quark-135m' },
|
| 593 |
+
{ name: 'Veyra-30M-Base', org: 'veyra-ai', params: 34611712, arc_easy: 0.3594, arc_challenge: 0.2389, piqa: 0.5903, hellaswag: 0.2787, avg_4: 0.3692, url: 'https://huggingface.co/veyra-ai/veyra-30m-base-5b-tokens' },
|
| 594 |
+
{ name: 'Syn-2.6M', org: 'FromZero', params: 2604210, arc_easy: 0.3211, arc_challenge: 0.2039, piqa: 0.5365, hellaswag: 0.2702, avg_4: 0.3331, url: 'https://huggingface.co/fromziro/Syn-2.6M' },
|
| 595 |
+
{ name: 'Ant-5M', org: 'GODELEV', params: 4713344, arc_easy: 0.2639, arc_challenge: 0.2577, piqa: 0.4864, hellaswag: 0.2599, avg_4: 0.3195, url: 'https://huggingface.co/GODELEV/Ant-5m' },
|
| 596 |
+
{ name: 'Er-Medium-12.5M', org: 'FromZero', params: 12497520, arc_easy: 0.3476, arc_challenge: 0.2082, piqa: 0.5783, hellaswag: 0.2841, avg_4: 0.3668, url: 'https://huggingface.co/fromziro/Er-Medium-12.5M' },
|
| 597 |
+
{ name: 'michel-tiny', org: 'finnianx', params: 55719040, arc_easy: 0.3737, arc_challenge: 0.2176, piqa: 0.5734, hellaswag: 0.2815, avg_4: 0.3502, url: 'https://huggingface.co/finnianx/michel-tiny' },
|
| 598 |
+
{ name: 'OdinNext-138M-Base', org: 'joelhenwang', params: 138449696, arc_easy: 0.4508, arc_challenge: 0.2381, piqa: 0.5952, hellaswag: 0.2809, avg_4: 0.3920, url: 'https://huggingface.co/joelhenwang/OdinNext-138M-Base' },
|
| 599 |
+
{ name: 'OdinNext-138M-Instruct', org: 'joelhenwang', params: 138451232, arc_easy: 0.4440, arc_challenge: 0.2312, piqa: 0.5865, hellaswag: 0.2886, avg_4: 0.3888, url: 'https://huggingface.co/joelhenwang/OdinNext-138M-Instruct' },
|
| 600 |
+
{ name: 'michel-micro', org: 'finnianx', params: 28355072, arc_easy: 0.3859, arc_challenge: 0.2312, piqa: 0.5762, hellaswag: 0.2823, avg_4: 0.3803, url: 'https://huggingface.co/finnianx/michel-micro' },
|
| 601 |
+
{ name: 'kirk-tung', org: 'RTC', params: 53111296, arc_easy: 0.3043, arc_challenge: 0.2201, piqa: 0.5261, hellaswag: 0.2632, avg_4: 0.3184, url: 'https://huggingface.co/rtc2022/kirk-tung' },
|
| 602 |
+
{ name: 'ettin-decoder-32m', org: 'JHU CLSP', params: 32081344, arc_easy: 0.3173, arc_challenge: 0.2201, piqa: 0.5229, hellaswag: 0.2779, avg_4: 0.3348, url: 'https://huggingface.co/jhu-clsp/ettin-decoder-32m' },
|
| 603 |
+
{ name: 'ettin-decoder-17m', org: 'JHU CLSP', params: 16913600, arc_easy: 0.3165, arc_challenge: 0.2227, piqa: 0.5337, hellaswag: 0.2739, avg_4: 0.3438, url: 'https://huggingface.co/jhu-clsp/ettin-decoder-17m' },
|
| 604 |
+
{ name: 'Er-Tiny-1.3M', org: 'FromZero', params: 1332744, arc_easy: 0.3009, arc_challenge: 0.2235, piqa: 0.5272, hellaswag: 0.2767, avg_4: 0.3292, url: 'https://huggingface.co/fromziro/Er-Tiny-1.3M' },
|
| 605 |
+
{ name: 'michel-nano', org: 'finnianx', params: 5963008, arc_easy: 0.3329, arc_challenge: 0.2270, piqa: 0.5479, hellaswag: 0.2712, avg_4: 0.3464, url: 'https://huggingface.co/finnianx/michel-nano' },
|
| 606 |
+
{ name: 'Supra-1.5-50M-base-exp', org: 'SupraLabs', params: 51786240, arc_easy: 0.4840, arc_challenge: 0.2551, piqa: 0.6001, hellaswag: 0.2978, avg_4: 0.4288, url: 'https://huggingface.co/SupraLabs/Supra-1.5-50M-base-exp' },
|
| 607 |
+
{ name: 'ettin-decoder-68m', org: 'JHU CLSP', params: 68457664, arc_easy: 0.3274, arc_challenge: 0.2304, piqa: 0.5218, hellaswag: 0.2781, avg_4: 0.3394, url: 'https://huggingface.co/jhu-clsp/ettin-decoder-68m' },
|
| 608 |
+
{ name: 'ettin-decoder-150m', org: 'JHU CLSP', params: 149655232, arc_easy: 0.2997, arc_challenge: 0.2423, piqa: 0.5125, hellaswag: 0.2595, avg_4: 0.3189, url: 'https://huggingface.co/jhu-clsp/ettin-decoder-150m' },
|
| 609 |
+
{ name: 'Ant-10M', org: 'GODELEV', params: 9902336, arc_easy: 0.2753, arc_challenge: 0.2824, piqa: 0.5065, hellaswag: 0.2696, avg_4: 0.3302, url: 'https://huggingface.co/GODELEV/Ant-19m' },
|
| 610 |
+
{ name: 'Supra-1.5-50M-Instruct-exp', org: 'SupraLabs', params: 51786240, arc_easy: 0.4394, arc_challenge: 0.2611, piqa: 0.5941, hellaswag: 0.2926, avg_4: 0.4077, url: 'https://huggingface.co/SupraLabs/Supra-1.5-50M-Instruct-exp' },
|
| 611 |
+
{ name: 'SLM-10M', org: 'Liodon AI', params: 9968640, arc_easy: 0.3552, arc_challenge: 0.2346, piqa: 0.5707, hellaswag: 0.2740, avg_4: 0.3401, url: 'https://huggingface.co/liodon-ai/slm-10m' },
|
| 612 |
+
{ name: 'Doge-20M', org: 'SmallDoge', params: 13118728, arc_easy: 0.3338, arc_challenge: 0.2073, piqa: 0.5430, hellaswag: 0.2661, avg_4: 0.3236, url: 'https://huggingface.co/SmallDoge/Doge-20M' }
|
| 613 |
];
|
| 614 |
|
| 615 |
const BENCHMARKS = [
|
| 616 |
+
{ key: 'avg_4', label: 'Avg' },
|
| 617 |
+
{ key: 'arc_easy', label: 'ARC-Easy' },
|
| 618 |
+
{ key: 'arc_challenge', label: 'ARC-Challenge' },
|
| 619 |
{ key: 'hellaswag', label: 'HellaSwag' },
|
| 620 |
+
{ key: 'piqa', label: 'PIQA' }
|
|
|
|
| 621 |
];
|
| 622 |
|
| 623 |
+
const BENCHMARK_NAMES = {
|
| 624 |
+
avg_4: 'Average',
|
| 625 |
+
arc_easy: 'ARC-Easy',
|
| 626 |
+
arc_challenge: 'ARC-Challenge',
|
| 627 |
+
hellaswag: 'HellaSwag',
|
| 628 |
+
piqa: 'PIQA'
|
| 629 |
+
};
|
| 630 |
+
|
| 631 |
+
let activeBenchmark = 'avg_4';
|
| 632 |
+
let activeSearch = '';
|
| 633 |
+
let sortKey = 'score';
|
| 634 |
let chart = null;
|
| 635 |
|
| 636 |
+
const fmtParams = (n) => {
|
| 637 |
if (n >= 1e9) return `${(n / 1e9).toFixed(2)}B`;
|
| 638 |
if (n >= 1e6) return `${(n / 1e6).toFixed(2)}M`;
|
| 639 |
if (n >= 1e3) return `${(n / 1e3).toFixed(2)}K`;
|
| 640 |
return String(n);
|
| 641 |
};
|
| 642 |
|
| 643 |
+
const toPct = (v) => (v * 100);
|
| 644 |
|
| 645 |
+
const linearRegression = (points) => {
|
|
|
|
|
|
|
| 646 |
const n = points.length;
|
| 647 |
let sumX = 0, sumY = 0, sumXY = 0, sumXX = 0;
|
| 648 |
for (const p of points) {
|
|
|
|
| 655 |
const slope = Math.abs(denom) < 1e-12 ? 0 : (n * sumXY - sumX * sumY) / denom;
|
| 656 |
const intercept = n ? (sumY - slope * sumX) / n : 0;
|
| 657 |
return { slope, intercept };
|
| 658 |
+
};
|
| 659 |
|
| 660 |
+
const getScore = (m, key) => (m[key] ?? null);
|
| 661 |
+
|
| 662 |
+
const getVisibleModels = () => {
|
| 663 |
+
const q = activeSearch.trim().toLowerCase();
|
| 664 |
+
return MODELS.filter(m => {
|
| 665 |
+
const s = getScore(m, activeBenchmark);
|
| 666 |
+
if (s === null || !Number.isFinite(m.params)) return false;
|
| 667 |
+
if (!q) return true;
|
| 668 |
+
return (
|
| 669 |
+
m.name.toLowerCase().includes(q) ||
|
| 670 |
+
m.org.toLowerCase().includes(q)
|
| 671 |
+
);
|
| 672 |
});
|
| 673 |
+
};
|
| 674 |
+
|
| 675 |
+
const computeLoss = (data, slope, intercept) => {
|
| 676 |
+
const residuals = data.map(d => d.y - (slope * d.x + intercept));
|
| 677 |
+
const mse = residuals.reduce((a, b) => a + b * b, 0) / residuals.length;
|
| 678 |
const rmse = Math.sqrt(mse);
|
| 679 |
return { mse, rmse, residuals };
|
| 680 |
+
};
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 681 |
|
| 682 |
+
const renderBenchmarkButtons = () => {
|
| 683 |
const box = document.getElementById('benchmarkButtons');
|
| 684 |
+
box.innerHTML = BENCHMARKS.map(b =>
|
| 685 |
+
`<button class="${b.key === activeBenchmark ? 'active' : ''}" data-key="${b.key}">${b.label}</button>`
|
| 686 |
+
).join('');
|
| 687 |
box.querySelectorAll('button').forEach(btn => {
|
| 688 |
btn.addEventListener('click', () => {
|
| 689 |
activeBenchmark = btn.dataset.key;
|
| 690 |
render();
|
| 691 |
});
|
| 692 |
});
|
| 693 |
+
};
|
| 694 |
|
| 695 |
+
const renderTable = (data) => {
|
| 696 |
+
const tbody = document.getElementById('modelTable');
|
| 697 |
+
const rows = [...data].sort((a, b) => {
|
| 698 |
+
if (sortKey === 'name') return a.name.localeCompare(b.name);
|
| 699 |
+
if (sortKey === 'org') return a.org.localeCompare(b.org);
|
| 700 |
+
if (sortKey === 'params') return a.params - b.params;
|
| 701 |
+
if (sortKey === 'score') return b.score - a.score;
|
| 702 |
+
if (sortKey === 'residual') return b.residual - a.residual;
|
| 703 |
+
if (sortKey === 'side') return b.residual - a.residual;
|
| 704 |
+
return 0;
|
| 705 |
+
});
|
|
|
|
| 706 |
|
| 707 |
+
tbody.innerHTML = rows.map(m => {
|
| 708 |
+
const above = m.residual >= 0;
|
| 709 |
+
return `
|
| 710 |
+
<tr data-url="${m.url}">
|
| 711 |
+
<td>${m.name}</td>
|
| 712 |
+
<td class="muted">${m.org}</td>
|
| 713 |
+
<td class="muted">${fmtParams(m.params)} <span style="color:#6f737d;">(${m.params.toLocaleString()})</span></td>
|
| 714 |
+
<td>${m.score.toFixed(2)}%</td>
|
| 715 |
+
<td class="${above ? 'above' : 'below'}">${above ? '+' : ''}${m.residual.toFixed(2)} pts</td>
|
| 716 |
+
<td>${above ? '<span class="pill-up">Above line</span>' : '<span class="pill-down">Below line</span>'}</td>
|
| 717 |
+
</tr>
|
| 718 |
+
`;
|
| 719 |
+
}).join('');
|
| 720 |
+
|
| 721 |
+
tbody.querySelectorAll('tr').forEach(row => {
|
| 722 |
row.addEventListener('click', () => {
|
| 723 |
const url = row.dataset.url;
|
| 724 |
if (url) window.open(url, '_blank', 'noopener,noreferrer');
|
| 725 |
});
|
| 726 |
});
|
| 727 |
+
};
|
| 728 |
|
| 729 |
+
const updateStats = (slope, intercept, mse, rmse, count) => {
|
| 730 |
document.getElementById('statSlope').textContent = slope.toFixed(4);
|
| 731 |
document.getElementById('statIntercept').textContent = intercept.toFixed(2);
|
| 732 |
+
document.getElementById('statMSE').textContent = mse.toFixed(4);
|
| 733 |
+
document.getElementById('statRMSE').textContent = rmse.toFixed(2);
|
| 734 |
document.getElementById('countBadge').textContent = `${count} models`;
|
| 735 |
+
document.getElementById('infoBenchmark').textContent = BENCHMARK_NAMES[activeBenchmark] || activeBenchmark;
|
| 736 |
+
document.getElementById('infoCount').textContent = String(count);
|
| 737 |
+
document.getElementById('infoLoss').textContent = `MSE ${mse.toFixed(4)} · RMSE ${rmse.toFixed(2)}`;
|
| 738 |
+
};
|
| 739 |
|
| 740 |
function render() {
|
| 741 |
renderBenchmarkButtons();
|
| 742 |
|
| 743 |
+
const data = getVisibleModels().map(m => ({
|
| 744 |
+
name: m.name,
|
| 745 |
+
org: m.org,
|
| 746 |
+
params: m.params,
|
| 747 |
+
score: toPct(m[activeBenchmark]),
|
| 748 |
+
url: m.url,
|
| 749 |
+
x: Math.log10(m.params)
|
| 750 |
+
}));
|
| 751 |
+
|
| 752 |
+
if (data.length < 2) {
|
| 753 |
+
document.getElementById('chartTitle').textContent = 'Not enough models';
|
| 754 |
+
document.getElementById('chartSub').textContent = 'Add more models or clear the search filter.';
|
| 755 |
+
document.getElementById('fitNote').textContent = 'Need at least 2 visible models to fit a line.';
|
| 756 |
+
document.getElementById('modelTable').innerHTML = '';
|
| 757 |
+
if (chart) chart.destroy();
|
| 758 |
+
return;
|
| 759 |
+
}
|
| 760 |
|
| 761 |
+
const points = data.map(d => ({ x: d.x, y: d.score }));
|
| 762 |
const { slope, intercept } = linearRegression(points);
|
| 763 |
+
const { mse, rmse, residuals } = computeLoss(data, slope, intercept);
|
| 764 |
+
|
| 765 |
+
data.forEach((d, i) => {
|
| 766 |
+
d.residual = residuals[i];
|
| 767 |
+
});
|
| 768 |
|
| 769 |
updateStats(slope, intercept, mse, rmse, data.length);
|
|
|
|
| 770 |
|
| 771 |
const xMin = Math.min(...data.map(d => d.x));
|
| 772 |
const xMax = Math.max(...data.map(d => d.x));
|
| 773 |
+
const xPad = Math.max(0.2, (xMax - xMin) * 0.09);
|
| 774 |
+
const xLow = xMin - xPad;
|
| 775 |
+
const xHigh = xMax + xPad;
|
| 776 |
const yLow = slope * xLow + intercept;
|
| 777 |
const yHigh = slope * xHigh + intercept;
|
| 778 |
+
|
| 779 |
+
const yMin = Math.min(...data.map(d => d.score), yLow) - 1.8;
|
| 780 |
+
const yMax = Math.max(...data.map(d => d.score), yHigh) + 1.8;
|
| 781 |
|
| 782 |
const titleMap = {
|
| 783 |
+
avg_4: 'Average score vs log parameters',
|
| 784 |
+
arc_easy: 'ARC-Easy vs log parameters',
|
| 785 |
+
arc_challenge: 'ARC-Challenge vs log parameters',
|
| 786 |
hellaswag: 'HellaSwag vs log parameters',
|
| 787 |
+
piqa: 'PIQA vs log parameters'
|
|
|
|
| 788 |
};
|
| 789 |
|
| 790 |
+
document.getElementById('chartTitle').textContent = titleMap[activeBenchmark] || 'Regression vs log parameters';
|
| 791 |
+
document.getElementById('chartSub').textContent = `Silver dotted regression line on ${BENCHMARK_NAMES[activeBenchmark] || activeBenchmark}. Hover for model names, click to open model pages.`;
|
| 792 |
+
document.getElementById('fitNote').textContent = `Regression loss for ${BENCHMARK_NAMES[activeBenchmark] || activeBenchmark}: MSE ${mse.toFixed(4)} · RMSE ${rmse.toFixed(2)}.`;
|
| 793 |
+
|
| 794 |
+
renderTable(data);
|
| 795 |
|
| 796 |
const ctx = document.getElementById('scatterChart').getContext('2d');
|
| 797 |
if (chart) chart.destroy();
|
|
|
|
| 808 |
{ x: xHigh, y: yHigh }
|
| 809 |
],
|
| 810 |
borderColor: 'rgba(236, 236, 241, 0.98)',
|
| 811 |
+
borderWidth: 3,
|
| 812 |
+
borderDash: [2, 7],
|
| 813 |
+
borderCapStyle: 'round',
|
| 814 |
pointRadius: 0,
|
| 815 |
tension: 0,
|
| 816 |
fill: false,
|
|
|
|
| 819 |
{
|
| 820 |
label: 'Models',
|
| 821 |
data: data.map(d => ({ ...d })),
|
| 822 |
+
parsing: false,
|
| 823 |
pointRadius: 7,
|
| 824 |
pointHoverRadius: 10,
|
| 825 |
+
pointHitRadius: 12,
|
| 826 |
borderWidth: 1.5,
|
| 827 |
+
backgroundColor: (ctx) => {
|
| 828 |
+
const raw = ctx.raw;
|
| 829 |
+
if (!raw) return 'rgba(225, 226, 232, 0.86)';
|
| 830 |
+
return raw.residual >= 0
|
| 831 |
+
? 'rgba(241, 242, 247, 0.96)'
|
| 832 |
+
: 'rgba(146, 151, 160, 0.93)';
|
| 833 |
+
},
|
| 834 |
+
borderColor: (ctx) => {
|
| 835 |
+
const raw = ctx.raw;
|
| 836 |
+
if (!raw) return 'rgba(255,255,255,0.92)';
|
| 837 |
+
return raw.residual >= 0
|
| 838 |
+
? 'rgba(255,255,255,0.98)'
|
| 839 |
+
: 'rgba(220,220,224,0.55)';
|
| 840 |
+
},
|
| 841 |
+
hoverBackgroundColor: 'rgba(247, 247, 250, 0.99)',
|
| 842 |
hoverBorderColor: 'rgba(255,255,255,0.98)',
|
| 843 |
order: 2
|
| 844 |
}
|
|
|
|
| 849 |
maintainAspectRatio: false,
|
| 850 |
animation: { duration: 0 },
|
| 851 |
interaction: { mode: 'nearest', intersect: true },
|
| 852 |
+
onHover: (event, elements) => {
|
| 853 |
+
event.native.target.style.cursor = elements.length ? 'pointer' : 'default';
|
| 854 |
+
},
|
| 855 |
onClick: (event, elements) => {
|
| 856 |
if (!elements.length) return;
|
| 857 |
const hit = elements[0];
|
|
|
|
| 862 |
plugins: {
|
| 863 |
legend: { display: false },
|
| 864 |
tooltip: {
|
| 865 |
+
backgroundColor: 'rgba(12,12,15,0.97)',
|
| 866 |
+
borderColor: 'rgba(236,236,241,0.24)',
|
| 867 |
borderWidth: 1,
|
| 868 |
titleColor: '#ffffff',
|
| 869 |
+
bodyColor: '#d7d7de',
|
| 870 |
padding: 12,
|
| 871 |
displayColors: false,
|
| 872 |
callbacks: {
|
|
|
|
| 876 |
const pred = slope * d.x + intercept;
|
| 877 |
const resid = d.y - pred;
|
| 878 |
return [
|
| 879 |
+
`Params: ${fmtParams(d.params)} (${d.params.toLocaleString()})`,
|
| 880 |
+
`Score: ${d.score.toFixed(2)}%`,
|
| 881 |
`Residual: ${resid >= 0 ? '+' : ''}${resid.toFixed(2)} pts`,
|
| 882 |
`Prediction: ${pred.toFixed(2)}%`
|
| 883 |
];
|
|
|
|
| 892 |
max: xHigh,
|
| 893 |
grid: { color: 'rgba(255,255,255,0.045)' },
|
| 894 |
ticks: {
|
| 895 |
+
color: '#84848d',
|
| 896 |
callback: v => `10^${Number(v).toFixed(1)}`
|
| 897 |
},
|
| 898 |
title: {
|
| 899 |
display: true,
|
| 900 |
text: 'Log₁₀(parameters)',
|
| 901 |
+
color: '#84848d',
|
| 902 |
font: { weight: '600' }
|
| 903 |
}
|
| 904 |
},
|
|
|
|
| 907 |
max: yMax,
|
| 908 |
grid: { color: 'rgba(255,255,255,0.045)' },
|
| 909 |
ticks: {
|
| 910 |
+
color: '#84848d',
|
| 911 |
callback: v => `${Number(v).toFixed(0)}%`
|
| 912 |
},
|
| 913 |
title: {
|
| 914 |
display: true,
|
| 915 |
text: 'Score (%)',
|
| 916 |
+
color: '#84848d',
|
| 917 |
font: { weight: '600' }
|
| 918 |
}
|
| 919 |
}
|
|
|
|
| 922 |
});
|
| 923 |
}
|
| 924 |
|
| 925 |
+
document.getElementById('searchBox').addEventListener('input', (e) => {
|
| 926 |
+
activeSearch = e.target.value || '';
|
| 927 |
+
render();
|
| 928 |
+
});
|
| 929 |
+
|
| 930 |
+
document.querySelectorAll('thead th[data-sort]').forEach(th => {
|
| 931 |
+
th.addEventListener('click', () => {
|
| 932 |
+
sortKey = th.dataset.sort;
|
| 933 |
+
render();
|
| 934 |
+
});
|
| 935 |
+
});
|
| 936 |
+
|
| 937 |
render();
|
| 938 |
</script>
|
| 939 |
</body>
|
| 940 |
+
</html>
|