Spaces:
Running
Running
File size: 24,080 Bytes
5a4fe61 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 | <!DOCTYPE html>
<html lang="en">
<head>
<meta charset="UTF-8">
<meta name="viewport" content="width=device-width, initial-scale=1.0">
<title>Supra 50M | SupraLabs Blog</title>
<style>
:root {
--bg: #0f0f0f;
--surface: #1a1a1a;
--border: #333;
--text: #e0e0e0;
--accent: #536bfe;
--muted: #888;
--font-mono: 'JetBrains Mono', 'Fira Code', monospace;
}
* { margin: 0; padding: 0; box-sizing: border-box; }
body {
background-color: var(--bg);
color: var(--text);
font-family: 'Inter', -apple-system, sans-serif;
line-height: 1.6;
padding: 2rem;
}
code, pre, .mono { font-family: var(--font-mono); }
.container { max-width: 900px; margin: 0 auto; }
header {
border-bottom: 2px solid var(--border);
padding-bottom: 2rem;
margin-bottom: 3rem;
display: flex;
justify-content: space-between;
align-items: flex-end;
}
.logo-area h1 {
font-size: 1.2rem;
text-transform: uppercase;
letter-spacing: 2px;
color: var(--accent);
line-height: 1;
display: flex;
align-items: center;
gap: 10px;
}
.logo-area a { text-decoration: none; color: inherit; }
.logo-area {
display: flex;
align-items: center;
gap: 10px;
font-weight: bold;
font-size: 1.2rem;
}
nav a {
color: var(--text);
text-decoration: none;
margin-left: 1.5rem;
font-size: 0.9rem;
border-bottom: 1px solid transparent;
}
nav a:hover { border-bottom: 1px solid var(--accent); }
.post-header { margin-bottom: 3rem; }
.post-header h2 {
font-size: 3rem;
line-height: 1.1;
margin-bottom: 1rem;
font-weight: 800;
}
.post-meta {
font-family: var(--font-mono);
color: var(--accent);
font-size: 0.9rem;
margin-bottom: 2rem;
}
.post-content {
background: var(--surface);
border: 1px solid var(--border);
padding: 3rem;
margin-bottom: 4rem;
}
.post-content h2 {
font-size: 1.8rem;
margin: 2.5rem 0 1rem 0;
color: var(--accent);
}
.post-content h2:first-child { margin-top: 0; }
.post-content p {
margin-bottom: 1.5rem;
font-size: 1.1rem;
color: var(--text);
}
.post-content ul {
margin-bottom: 1.5rem;
padding-left: 1.5rem;
}
.post-content li { margin-bottom: 0.5rem; font-size: 1.1rem; }
.post-content strong { color: #fff; }
.post-content code {
background: #111;
border: 1px solid var(--border);
padding: 2px 6px;
border-radius: 3px;
font-size: 0.95em;
color: var(--accent);
}
.code-block {
background: #111;
border: 1px solid var(--border);
padding: 1.5rem;
margin: 2rem 0;
overflow-x: auto;
font-family: var(--font-mono);
font-size: 0.85rem;
line-height: 1.7;
color: #ccc;
}
.code-block .comment { color: var(--muted); }
.code-block .keyword { color: var(--accent); }
.code-block .string { color: #a8d8a8; }
.callout {
border-left: 3px solid var(--accent);
background: #111;
padding: 1rem 1.5rem;
margin: 2rem 0;
font-family: var(--font-mono);
font-size: 0.95rem;
color: #ccc;
}
.callout span {
display: block;
color: var(--muted);
font-size: 0.8rem;
margin-bottom: 0.4rem;
}
.output-example {
border: 1px solid var(--border);
background: #111;
padding: 1.5rem;
margin: 1.5rem 0;
}
.output-example .prompt-label {
font-family: var(--font-mono);
font-size: 0.75rem;
color: var(--accent);
margin-bottom: 0.4rem;
}
.output-example .prompt-text {
font-weight: 700;
color: #fff;
margin-bottom: 1rem;
font-size: 1rem;
}
.output-example .output-label {
font-family: var(--font-mono);
font-size: 0.75rem;
color: var(--muted);
margin-bottom: 0.4rem;
}
.output-example .output-text {
color: var(--text);
font-size: 0.95rem;
font-style: italic;
line-height: 1.7;
}
/* Badge para Base / Instruct */
.model-badge {
display: inline-block;
font-family: var(--font-mono);
font-size: 0.72rem;
padding: 3px 10px;
border-radius: 3px;
margin-right: 0.5rem;
vertical-align: middle;
}
.badge-base {
background: #1a1a2e;
border: 1px solid var(--accent);
color: var(--accent);
}
.badge-instruct {
background: #1a2e1a;
border: 1px solid #5bfe7a;
color: #5bfe7a;
}
.table-wrap { overflow-x: auto; margin: 2rem 0; }
table {
width: 100%;
border-collapse: collapse;
font-family: var(--font-mono);
font-size: 0.88rem;
}
th {
background: #111;
color: var(--accent);
padding: 0.75rem 1rem;
text-align: left;
border: 1px solid var(--border);
}
td {
padding: 0.7rem 1rem;
border: 1px solid var(--border);
color: var(--text);
}
tr:nth-child(even) td { background: #111; }
.highlight-row td { color: #fff; font-weight: 600; }
.tags { display: flex; gap: 0.5rem; margin-top: 2rem; flex-wrap: wrap; }
.tag {
font-family: var(--font-mono);
font-size: 0.7rem;
padding: 2px 8px;
border: 1px solid var(--border);
border-radius: 4px;
color: var(--muted);
}
footer {
margin-top: 6rem;
padding-bottom: 2rem;
font-size: 0.8rem;
color: var(--muted);
text-align: center;
}
@media (max-width: 600px) {
.post-header h2 { font-size: 2rem; }
.post-content { padding: 1.5rem; }
header { flex-direction: column; align-items: flex-start; gap: 1rem; }
nav a { margin-left: 0; margin-right: 1rem; }
}
</style>
</head>
<body>
<div class="container">
<header>
<div class="logo-area" style="font-size: 1.5em;">
<a href="./index.html"><h1><img src="./image.png" style="height: 2em"> SupraLabs_</h1></a>
</div>
<nav>
<a href="./index.html#news">News</a>
<a href="https://huggingface.co/SupraLabs" target="blank">HuggingFace</a>
<a href="./index.html#hardware">Hardware</a>
</nav>
</header>
<article>
<div class="post-header">
<div class="post-meta">// 2026-05-21 | Release</div>
<h2>🦅 Supra 50M:<br>Base and Instruct,<br>both live today.</h2>
</div>
<div class="post-content">
<p>Today is a big day for us. We are shipping two models at once: <strong>Supra-50M Base</strong> and <strong>Supra-50M Instruct</strong>. This is the first release under our <strong>Scaling Up Plan</strong>, and the jump from 8M to 50M parameters is the largest single leap we have made so far. The Base model was pretrained from scratch on 20 billion tokens. The Instruct model was then fine-tuned on top of it for chat and instruction following. Two very different use cases, one release.</p>
<h2>Supra-50M Base <span class="model-badge badge-base">BASE</span></h2>
<p>The Base model is a raw pretraining run. No instructions, no chat format, just a Llama-architecture decoder trained to predict the next token on high-quality educational text. The full config:</p>
<div class="callout">
<span>// supra-50m base — model config</span>
Parameters → ~50M (51.8M)<br>
Architecture → Llama (decoder-only)<br>
Vocab size → 32,000 (custom BPE)<br>
Hidden size → 512<br>
Intermediate size → 1,408<br>
Layers → 12<br>
Attention heads → 8 (GQA: 4 KV heads)<br>
Context length → 1,024 tokens<br>
Training tokens → 20,000,000,000 (20B)<br>
Dataset → Fineweb-Edu (sample-100BT)<br>
Final loss → 3.259<br>
Precision → bfloat16
</div>
<p>The tokenizer was trained from scratch on 500,000 documents from Fineweb-Edu using Byte-Level BPE with a 32,000 vocabulary. Training ran on a single GPU with <code>torch.compile</code> enabled, cosine LR schedule, and AdamW fused optimizer.</p>
<h2>Benchmarks</h2>
<p>We benchmarked the Base model against GPT-2 (124M), SmolLM-135M, and OpenELM-270M. Supra-50M uses 2 to 5 times fewer parameters than every model in the comparison.</p>
<div class="table-wrap">
<table>
<thead>
<tr>
<th>Benchmark</th>
<th>Supra-50M</th>
<th>GPT-2 (124M)</th>
<th>SmolLM (135M)</th>
<th>OpenELM (270M)</th>
</tr>
</thead>
<tbody>
<tr>
<td>Parameters</td>
<td><strong>50M</strong></td>
<td>124M (2.5×)</td>
<td>135M (2.7×)</td>
<td>270M (5.4×)</td>
</tr>
<tr class="highlight-row">
<td>BLiMP (linguistics)</td>
<td>76.3%</td>
<td>63.0%</td>
<td>69.8%</td>
<td>N/A</td>
</tr>
<tr>
<td>SciQ (science)</td>
<td>77.2%</td>
<td>53.2%</td>
<td>73.4%</td>
<td>84.7%</td>
</tr>
<tr class="highlight-row">
<td>ARC-Easy (knowledge)</td>
<td>52.2%</td>
<td>42.0%</td>
<td>49.2%</td>
<td>45.1%</td>
</tr>
<tr>
<td>PIQA (logic)</td>
<td>62.2%</td>
<td>63.0%</td>
<td>67.3%</td>
<td>69.8%</td>
</tr>
<tr>
<td>HellaSwag (context)</td>
<td>31.8%</td>
<td>29.5%</td>
<td>42.0%</td>
<td>46.7%</td>
</tr>
</tbody>
</table>
</div>
<p>BLiMP at 76.3% beats GPT-2 by over 13 points and SmolLM by 6.5 points, with a model that is 2.5 to 2.7 times smaller. ARC-Easy at 52.2% beats all three comparison models including those much larger. The weaker numbers on HellaSwag and PIQA reflect the longer-context reasoning those tasks require, which is an area to push on next.</p>
<h2>Supra-50M Instruct <span class="model-badge badge-instruct">INSTRUCT</span></h2>
<p>The Instruct model takes the Base weights and fine-tunes them for instruction following using the full <strong>Alpaca-Cleaned</strong> dataset for 4 epochs. The prompt format used at training time, which you need to match at inference time, looks like this:</p>
<div class="callout">
<span>// instruct prompt template</span>
Below is an instruction that describes a task. Write a response<br>
that appropriately completes the request.<br>
<br>
### Instruction:<br>
{your instruction here}<br>
<br>
### Response:<br>
{model output starts here}
</div>
<p>If you send a raw prompt to the Instruct model without this format, you will get Base-model behavior. The template matters.</p>
<h2>Base model outputs</h2>
<div class="output-example">
<div class="prompt-label">// base model · prompt</div>
<div class="prompt-text">"The main concept of physics is "</div>
<div class="output-label">// output</div>
<div class="output-text">The main concept of physics is iffy, and the idea that we can make things behave in a certain way. The most important part of physics is called quantum mechanics which states that all particles are made up of energy (energy) and matter (matter). In physics, there are two types of particles: elementary particles and exotic ones...</div>
</div>
<div class="output-example">
<div class="prompt-label">// base model · prompt</div>
<div class="prompt-text">"Once upon a time, "</div>
<div class="output-label">// output</div>
<div class="output-text">Once upon a time, I was so excited about the new school year and wanted to make some changes in my life. I had been looking for ways to help me become more self-aware. As an adult, I have always felt that there is no one way of doing things without thinking first...</div>
</div>
<h2>Instruct model output</h2>
<div class="output-example">
<div class="prompt-label">// instruct model · user</div>
<div class="prompt-text">What is AI?</div>
<div class="output-label">// response</div>
<div class="output-text">AI refers to the simulation of human intelligence and creativity in machines, which are designed to mimic or surpass human cognition and reasoning. It is the process by which AI systems can solve problems, make predictions, or automate processes that typically require human-level thinking and problem solving. AI technology has been rapidly growing and evolving over time, with many potential benefits for society as a whole...</div>
</div>
<p>The difference between Base and Instruct outputs is clear. The Base model completes text naturally. The Instruct model answers questions with structure, numbered lists, and a coherent response format. That is exactly what SFT is supposed to do.</p>
<h2>How to run</h2>
<p><span class="model-badge badge-base">BASE</span> Standard pipeline, same pattern as previous Supra Mini releases:</p>
<div class="code-block">
<span class="keyword">from</span> transformers <span class="keyword">import</span> pipeline<br>
<span class="keyword">import</span> torch<br>
<br>
<span class="keyword">print</span>(<span class="string">"[*] Loading Supra-50M Base..."</span>)<br>
pipe = pipeline(<br>
<span class="string">"text-generation"</span>,<br>
model=<span class="string">"SupraLabs/Supra-50M-Base"</span>,<br>
device_map=<span class="string">"auto"</span>,<br>
torch_dtype=torch.float16 <span class="keyword">if</span> torch.cuda.is_available() <span class="keyword">else</span> torch.float32<br>
)<br>
<br>
<span class="keyword">def</span> generate_text(prompt, max_new_tokens=150):<br>
result = pipe(<br>
prompt,<br>
max_new_tokens=max_new_tokens,<br>
do_sample=True,<br>
temperature=0.5,<br>
top_k=25,<br>
top_p=0.9,<br>
repetition_penalty=1.2,<br>
pad_token_id=pipe.tokenizer.pad_token_id,<br>
eos_token_id=pipe.tokenizer.eos_token_id<br>
)<br>
<span class="keyword">return</span> result[0][<span class="string">'generated_text'</span>]<br>
<br>
<span class="keyword">print</span>(generate_text(<span class="string">"The importance of education is"</span>))
</div>
<p><span class="model-badge badge-instruct">INSTRUCT</span> The Instruct model needs the prompt template and a slightly different setup:</p>
<div class="code-block">
<span class="keyword">import</span> os, warnings<br>
os.environ[<span class="string">"TF_CPP_MIN_LOG_LEVEL"</span>] = <span class="string">"3"</span><br>
warnings.filterwarnings(<span class="string">"ignore"</span>, category=UserWarning, module=<span class="string">"transformers"</span>)<br>
<br>
<span class="keyword">import</span> torch<br>
<span class="keyword">from</span> transformers <span class="keyword">import</span> pipeline, AutoTokenizer, logging<br>
logging.set_verbosity_error()<br>
<br>
MODEL_ID = <span class="string">"SupraLabs/Supra-50M-Instruct"</span><br>
tokenizer = AutoTokenizer.from_pretrained(MODEL_ID, clean_up_tokenization_spaces=False)<br>
pipe = pipeline(<br>
<span class="string">"text-generation"</span>,<br>
model=MODEL_ID,<br>
tokenizer=tokenizer,<br>
device_map=<span class="string">"auto"</span>,<br>
torch_dtype=torch.bfloat16 <span class="keyword">if</span> torch.cuda.is_available() <span class="keyword">else</span> torch.float32<br>
)<br>
<br>
<span class="keyword">def</span> build_prompt(instruction, input_text=<span class="string">""</span>):<br>
<span class="keyword">if</span> input_text.strip():<br>
<span class="keyword">return</span> (<span class="string">"Below is an instruction that describes a task, paired with an input "<br>
"that provides further context. Write a response that appropriately "<br>
"completes the request.\n\n"</span><br>
f<span class="string">"### Instruction:\n{instruction}\n\n"</span><br>
f<span class="string">"### Input:\n{input_text}\n\n### Response:\n"</span>)<br>
<span class="keyword">return</span> (<span class="string">"Below is an instruction that describes a task. Write a response that "<br>
"appropriately completes the request.\n\n"</span><br>
f<span class="string">"### Instruction:\n{instruction}\n\n### Response:\n"</span>)<br>
<br>
<span class="keyword">def</span> generate(instruction, input_text=<span class="string">""</span>):<br>
result = pipe(build_prompt(instruction, input_text),<br>
max_new_tokens=512, do_sample=True, temperature=0.7,<br>
top_k=50, top_p=0.9, repetition_penalty=1.15,<br>
pad_token_id=pipe.tokenizer.pad_token_id,<br>
eos_token_id=pipe.tokenizer.eos_token_id,<br>
return_full_text=False)<br>
<span class="keyword">return</span> result[0][<span class="string">'generated_text'</span>].strip()<br>
<br>
<span class="keyword">print</span>(generate(<span class="string">"Explain what a neural network is."</span>))
</div>
<h2>What comes next</h2>
<p>Supra-50M is the first step of the Scaling Up Plan. The Base model showed that our training setup and custom tokenizer scale well, and the Instruct fine-tune showed that the Base weights are a solid foundation for supervised learning. The next goals are pushing past 50M, experimenting with longer context, and improving HellaSwag and PIQA numbers where we are still behind models with more parameters.</p>
<p><strong>Both models are live on HuggingFace right now. Go build something.</strong></p>
<div class="callout">
<span>// links</span>
Base → huggingface.co/SupraLabs/Supra-50M-Base<br>
Instruct → huggingface.co/SupraLabs/Supra-50M-Instruct<br>
License → Apache 2.0<br>
Series → Project Chimera / Supra collection on HuggingFace
</div>
<div class="tags">
<span class="tag">#release</span>
<span class="tag">#supra-50m</span>
<span class="tag">#chimera</span>
<span class="tag">#base-model</span>
<span class="tag">#instruct</span>
<span class="tag">#sft</span>
<span class="tag">#tinyml</span>
<span class="tag">#llama</span>
<span class="tag">#open-source</span>
<span class="tag">#fineweb-edu</span>
</div>
</div>
</article>
<footer>
<p class="mono">© 2026 SupraLabs // Built for the community.</p>
</footer>
</div>
</body>
</html> |