File size: 65,466 Bytes
40c4caa e7cc9a6 40c4caa | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 | <!DOCTYPE html>
<html lang="en">
<head>
<meta charset="UTF-8">
<meta name="viewport" content="width=device-width, initial-scale=1.0">
<title>LLM Pretraining Pipeline - 100M Params, Single GPU</title>
<style>
* { margin: 0; padding: 0; box-sizing: border-box; }
body {
font-family: 'Segoe UI', system-ui, -apple-system, sans-serif;
background: #0f172a;
color: #e2e8f0;
line-height: 1.7;
padding: 20px;
}
.container { max-width: 1100px; margin: 0 auto; }
h1 {
color: #38bdf8;
font-size: 2.2rem;
margin-bottom: 10px;
border-bottom: 2px solid #38bdf8;
padding-bottom: 15px;
}
h2 {
color: #7dd3fc;
font-size: 1.5rem;
margin: 30px 0 15px;
display: flex;
align-items: center;
gap: 10px;
}
h2::before {
content: "▸";
color: #38bdf8;
}
h3 {
color: #bae6fd;
font-size: 1.2rem;
margin: 20px 0 10px;
}
.subtitle {
color: #94a3b8;
font-size: 1.1rem;
margin-bottom: 25px;
}
.badge {
display: inline-block;
background: #1e293b;
border: 1px solid #334155;
padding: 4px 12px;
border-radius: 20px;
font-size: 0.85rem;
color: #94a3b8;
margin-right: 8px;
margin-bottom: 8px;
}
.badge.highlight {
background: #38bdf8;
color: #0f172a;
border-color: #38bdf8;
font-weight: 600;
}
.code-block {
background: #1e293b;
border: 1px solid #334155;
border-radius: 10px;
padding: 20px;
margin: 15px 0;
overflow-x: auto;
position: relative;
}
.code-block .lang {
position: absolute;
top: 8px;
right: 12px;
font-size: 0.75rem;
color: #64748b;
text-transform: uppercase;
}
pre {
font-family: 'Fira Code', 'Cascadia Code', 'JetBrains Mono', monospace;
font-size: 0.9rem;
line-height: 1.6;
color: #e2e8f0;
}
.comment { color: #6b7280; font-style: italic; }
.string { color: #a5f3fc; }
.keyword { color: #c084fc; }
.function { color: #38bdf8; }
.number { color: #fbbf24; }
.variable { color: #fca5a5; }
.note {
background: rgba(56, 189, 248, 0.1);
border-left: 4px solid #38bdf8;
padding: 15px 20px;
margin: 20px 0;
border-radius: 0 8px 8px 0;
}
.warning {
background: rgba(251, 191, 36, 0.1);
border-left: 4px solid #fbbf24;
}
.success {
background: rgba(52, 211, 153, 0.1);
border-left: 4px solid #34d399;
}
table {
width: 100%;
border-collapse: collapse;
margin: 15px 0;
font-size: 0.95rem;
}
th, td {
padding: 12px 15px;
text-align: left;
border-bottom: 1px solid #334155;
}
th {
color: #38bdf8;
font-weight: 600;
background: #1e293b;
}
tr:hover { background: rgba(30, 41, 59, 0.5); }
.step {
background: #1e293b;
border-radius: 12px;
padding: 20px;
margin: 15px 0;
border: 1px solid #334155;
}
.step-number {
display: inline-flex;
align-items: center;
justify-content: center;
width: 32px;
height: 32px;
background: #38bdf8;
color: #0f172a;
border-radius: 50%;
font-weight: bold;
margin-right: 10px;
font-size: 0.9rem;
}
.toc {
background: #1e293b;
border-radius: 12px;
padding: 20px;
margin: 20px 0;
border: 1px solid #334155;
}
.toc a {
color: #94a3b8;
text-decoration: none;
display: block;
padding: 6px 0;
border-bottom: 1px dotted #334155;
}
.toc a:hover { color: #38bdf8; }
.toc a:last-child { border-bottom: none; }
.gpu-info {
display: grid;
grid-template-columns: repeat(auto-fit, minmax(250px, 1fr));
gap: 15px;
margin: 20px 0;
}
.gpu-card {
background: #1e293b;
border-radius: 10px;
padding: 15px;
border: 1px solid #334155;
}
.gpu-card h4 { color: #38bdf8; margin-bottom: 8px; }
.progress-bar {
height: 8px;
background: #334155;
border-radius: 4px;
overflow: hidden;
margin-top: 8px;
}
.progress-fill {
height: 100%;
background: linear-gradient(90deg, #38bdf8, #7dd3fc);
border-radius: 4px;
}
.command {
background: #0f172a;
padding: 3px 8px;
border-radius: 4px;
font-family: monospace;
color: #fbbf24;
}
ul, ol { margin-left: 25px; margin-bottom: 15px; }
li { margin-bottom: 8px; }
a { color: #38bdf8; }
a:hover { color: #7dd3fc; }
.footer {
text-align: center;
margin-top: 40px;
padding: 20px;
color: #64748b;
border-top: 1px solid #334155;
}
.file-tree {
font-family: 'Fira Code', monospace;
background: #1e293b;
padding: 15px 20px;
border-radius: 8px;
border: 1px solid #334155;
line-height: 2;
}
.file-tree .dir { color: #38bdf8; }
.file-tree .file { color: #94a3b8; }
.file-tree .comment-tree { color: #6b7280; }
</style>
</head>
<body>
<div class="container">
<h1>🧠 Complete LLM Pretraining Pipeline</h1>
<p class="subtitle">Build a ~100M parameter general-text language model from scratch on a single GPU</p>
<div>
<span class="badge highlight">100M Parameters</span>
<span class="badge highlight">Single GPU</span>
<span class="badge highlight">General Text</span>
<span class="badge">Causal LM (GPT-style)</span>
<span class="badge">Hugging Face Ecosystem</span>
<span class="badge">Open Datasets</span>
</div>
<div class="toc">
<h3 style="margin-top:0;color:#38bdf8;">📋 Table of Contents</h3>
<a href="#overview">1. Project Overview & Architecture</a>
<a href="#setup">2. Environment Setup</a>
<a href="#data">3. Data Sourcing & Preparation</a>
<a href="#tokenizer">4. Train Custom Tokenizer</a>
<a href="#model">5. Model Configuration</a>
<a href="#training">6. Training Script</a>
<a href="#optimization">7. Memory Optimizations</a>
<a href="#evaluation">8. Evaluation & Benchmarking</a>
<a href="#inference">9. Inference & Text Generation</a>
<a href="#alignment">10. Optional: SFT/DPO Alignment</a>
</div>
<!-- SECTION 1: OVERVIEW -->
<h2 id="overview">1. Project Overview & Architecture</h2>
<div class="note">
<strong>Goal:</strong> Train a small but capable causal language model from scratch using only publicly available data and a single GPU (8-24GB VRAM).
</div>
<h3>Model Architecture (100M params)</h3>
<table>
<tr><th>Component</th><th>Setting</th><th>Notes</th></tr>
<tr><td>Architecture</td><td>GPT-Neo style (decoder-only)</td><td>Proven, efficient, well-supported</td></tr>
<tr><td>Hidden Size</td><td>768</td><td>Embedding dimension</td></tr>
<tr><td>Layers</td><td>12</td><td>Transformer blocks</td></tr>
<tr><td>Attention Heads</td><td>12</td><td>64 dim per head</td></tr>
<tr><td>FFN Dim</td><td>3072</td><td>4× hidden size</td></tr>
<tr><td>Vocab Size</td><td>32,000</td><td>Trained BPE tokenizer</td></tr>
<tr><td>Context Length</td><td>512</td><td>Trainable on single GPU</td></tr>
<tr><td>Parameters</td><td>~100M</td><td>Perfect for experimentation</td></tr>
</table>
<h3>Recommended GPU Specs</h3>
<div class="gpu-info">
<div class="gpu-card">
<h4>💻 Minimum: RTX 3060 12GB</h4>
<p>Batch size 4-8, gradient accumulation</p>
<div class="progress-bar"><div class="progress-fill" style="width:60%"></div></div>
</div>
<div class="gpu-card">
<h4>🎯 Recommended: RTX 4090 24GB</h4>
<p>Batch size 16-32, faster training</p>
<div class="progress-bar"><div class="progress-fill" style="width:90%"></div></div>
</div>
<div class="gpu-card">
<h4>☁️ Cloud: A100 40GB / L4 24GB</h4>
<p>Maximum throughput, bf16 support</p>
<div class="progress-bar"><div class="progress-fill" style="width:100%"></div></div>
</div>
</div>
<div class="note warning">
<strong>Training Time Estimate:</strong> On RTX 4090, ~2-4 hours for 1 epoch over 10B tokens. Full convergence may take 3-10 epochs depending on data quality.
</div>
<!-- SECTION 2: SETUP -->
<h2 id="setup">2. Environment Setup</h2>
<div class="step">
<span class="step-number">1</span>
<strong>Create project structure:</strong>
</div>
<div class="file-tree">
<span class="dir">llm-project/</span>
├── <span class="dir">data/</span> <span class="comment-tree"># Raw and processed datasets</span>
├── <span class="dir">tokenizer/</span> <span class="comment-tree"># Trained tokenizer files</span>
├── <span class="dir">models/</span> <span class="comment-tree"># Saved model checkpoints</span>
├── <span class="dir">scripts/</span> <span class="comment-tree"># Training scripts</span>
│ ├── <span class="file">train_tokenizer.py</span>
│ ├── <span class="file">pretrain.py</span>
│ ├── <span class="file">evaluate.py</span>
│ └── <span class="file">generate.py</span>
├── <span class="file">requirements.txt</span>
└── <span class="file">README.md</span>
</div>
<div class="step">
<span class="step-number">2</span>
<strong>Install dependencies:</strong>
</div>
<div class="code-block">
<span class="lang">bash</span>
<pre>
<span class="comment"># Create virtual environment</span>
python -m venv venv
source venv/bin/activate <span class="comment"># Linux/Mac</span>
<span class="comment"># venv\Scripts\activate # Windows</span>
<span class="comment"># Install core dependencies</span>
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121
pip install transformers datasets tokenizers accelerate wandb
pip install huggingface_hub sentencepiece protobuf
<span class="comment"># Optional: for optimization</span>
pip install flash-attn --no-build-isolation <span class="comment"># Requires CUDA 11.6+</span>
pip install bitsandbytes deepspeed
<span class="comment"># For evaluation</span>
pip install lm-eval
<span class="comment"># For alignment (SFT/DPO)</span>
pip install trl peft
</pre>
</div>
<div class="code-block">
<span class="lang">txt</span>
<span class="comment"># requirements.txt (save this)</span>
<pre>
torch>=2.1.0
transformers>=4.40.0
datasets>=2.18.0
tokenizers>=0.19.0
accelerate>=0.29.0
huggingface-hub>=0.22.0
wandb>=0.16.0
sentencepiece>=0.2.0
protobuf>=4.25.0
flash-attn>=2.5.0
bitsandbytes>=0.43.0
deepspeed>=0.14.0
trl>=0.8.0
peft>=0.10.0
lm-eval>=0.4.0
</pre>
</div>
<div class="note">
<strong>Login to Hugging Face:</strong> You'll need a token to upload models and access some datasets.
<div class="code-block" style="margin-top:10px;padding:10px;">
<pre>
huggingface-cli login
<span class="comment"># Or in Python:</span>
<span class="keyword">from</span> huggingface_hub <span class="keyword">import</span> login
login(token=<span class="string">"your_hf_token"</span>)
</pre>
</div>
</div>
<!-- SECTION 3: DATA -->
<h2 id="data">3. Data Sourcing & Preparation</h2>
<div class="note success">
We'll use a mix of <strong>high-quality open datasets</strong> that together provide ~15-20B tokens — perfect for training a 100M model.
</div>
<h3>Recommended Dataset Mix</h3>
<table>
<tr><th>Dataset</th><th>Size</th><th>Type</th><th>HF Hub ID</th></tr>
<tr><td>Fineweb-Edu</td><td>~1.3B tokens</td><td>Educational web text</td><td><code>HuggingFaceFW/fineweb-edu</code></td></tr>
<tr><td>The Pile (subset)</td><td>~2B tokens</td><td>Diverse academic/text</td><td><code>EleutherAI/pile</code></td></tr>
<tr><td>Project Gutenberg (PG19)</td><td>~3B tokens</td><td>Classic literature</td><td><code>pg19</code></td></tr>
<tr><td>OpenWebText</td><td>~8B tokens</td><td>Web text (Reddit links)</td><td><code>Skylion007/openwebtext</code></td></tr>
<tr><td>C4 (en)</td><td>~15B tokens</td><td>Cleaned Common Crawl</td><td><code>allenai/c4</code></td></tr>
</table>
<div class="step">
<span class="step-number">3</span>
<strong>Create data preparation script:</strong>
</div>
<div class="code-block">
<span class="lang">python</span>
<pre><span class="comment"># scripts/prepare_data.py</span>
<span class="keyword">import</span> os
<span class="keyword">from</span> datasets <span class="keyword">import</span> load_dataset, concatenate_datasets, DatasetDict
<span class="keyword">from</span> transformers <span class="keyword">import</span> AutoTokenizer
<span class="keyword">import</span> random
<span class="variable">SEED</span> = <span class="number">42</span>
random.seed(<span class="variable">SEED</span>)
<span class="keyword">def</span> <span class="function">load_and_mix_datasets</span>(target_tokens=<span class="number">5_000_000_000</span>, streaming=<span class="keyword">True</span>):
<span class="string">"""
Load and mix datasets for pretraining.
Target: ~5B tokens (good starting point for 100M model)
"""</span>
datasets_config = [
{<span class="string">"name"</span>: <span class="string">"HuggingFaceFW/fineweb-edu"</span>, <span class="string">"split"</span>: <span class="string">"train"</span>, <span class="string">"weight"</span>: <span class="number">0.30</span>, <span class="string">"text_col"</span>: <span class="string">"text"</span>},
{<span class="string">"name"</span>: <span class="string">"Skylion007/openwebtext"</span>, <span class="string">"split"</span>: <span class="string">"train"</span>, <span class="string">"weight"</span>: <span class="number">0.25</span>, <span class="string">"text_col"</span>: <span class="string">"text"</span>},
{<span class="string">"name"</span>: <span class="string">"allenai/c4"</span>, <span class="string">"split"</span>: <span class="string">"train"</span>, <span class="string">"weight"</span>: <span class="number">0.25</span>, <span class="string">"text_col"</span>: <span class="string">"text"</span>},
{<span class="string">"name"</span>: <span class="string">"bookcorpus"</span>, <span class="string">"split"</span>: <span class="string">"train"</span>, <span class="string">"weight"</span>: <span class="number">0.20</span>, <span class="string">"text_col"</span>: <span class="string">"text"</span>},
]
all_datasets = []
<span class="keyword">for</span> config <span class="keyword">in</span> datasets_config:
print(<span class="string">f"Loading {config['name']}..."</span>)
<span class="keyword">try</span>:
ds = load_dataset(
config[<span class="string">"name"</span>],
split=config[<span class="string">"split"</span>],
streaming=streaming,
trust_remote_code=<span class="keyword">True</span>
)
<span class="comment"># Take weighted subset</span>
ds = ds.shuffle(seed=<span class="variable">SEED</span>)
all_datasets.append((ds, config[<span class="string">"text_col"</span>], config[<span class="string">"weight"</span>]))
<span class="keyword">except</span> Exception <span class="keyword">as</span> e:
print(<span class="string">f"Warning: Could not load {config['name']}: {e}"</span>)
<span class="keyword">return</span> all_datasets
<span class="keyword">def</span> <span class="function">tokenize_function</span>(examples, tokenizer, text_column=<span class="string">"text"</span>):
<span class="string">"""Tokenize texts with truncation and padding."""</span>
<span class="keyword">return</span> tokenizer(
examples[text_column],
truncation=<span class="keyword">True</span>,
max_length=<span class="number">512</span>,
return_overflowing_tokens=<span class="keyword">False</span>,
return_length=<span class="keyword">True</span>,
)
<span class="keyword">def</span> <span class="function">prepare_pretraining_data</span>(
tokenizer_name_or_path=<span class="string">"gpt2"</span>,
output_dir=<span class="string">"./data/pretrain"</span>,
context_length=<span class="number">512</span>,
num_proc=<span class="number">4</span>
):
<span class="string">"""
Prepare concatenated pretraining data.
Groups texts into context-length chunks for efficient training.
"""</span>
os.makedirs(output_dir, exist_ok=<span class="keyword">True</span>)
<span class="comment"># Load tokenizer (we'll train our own later, use GPT-2 for now)</span>
tokenizer = AutoTokenizer.from_pretrained(tokenizer_name_or_path)
<span class="keyword">if</span> tokenizer.pad_token <span class="keyword">is</span> <span class="keyword">None</span>:
tokenizer.pad_token = tokenizer.eos_token
print(<span class="string">"Loading datasets..."</span>)
<span class="comment"># For initial prep, use a single reliable dataset</span>
ds = load_dataset(<span class="string">"openwebtext"</span>, streaming=<span class="keyword">True</span>, split=<span class="string">"train"</span>)
ds = ds.take(<span class="number">5_000_000</span>) <span class="comment"># ~2.5B tokens worth</span>
<span class="comment"># Convert streaming to in-memory for processing</span>
print(<span class="string">"Converting to in-memory dataset..."</span>)
texts = []
<span class="keyword">for</span> i, example <span class="keyword">in</span> enumerate(ds):
texts.append(example[<span class="string">"text"</span>])
<span class="keyword">if</span> i % <span class="number">100000</span> == <span class="number">0</span>:
print(<span class="string">f" Processed {i} examples..."</span>)
<span class="keyword">from</span> datasets <span class="keyword">import</span> Dataset
ds = Dataset.from_dict({<span class="string">"text"</span>: texts})
<span class="comment"># Tokenize</span>
print(<span class="string">"Tokenizing..."</span>)
<span class="keyword">def</span> <span class="function">tokenize_batch</span>(batch):
<span class="keyword">return</span> tokenizer(batch[<span class="string">"text"</span>], truncation=<span class="keyword">True</span>, max_length=context_length)
tokenized = ds.map(tokenize_batch, batched=<span class="keyword">True</span>, num_proc=num_proc, remove_columns=ds.column_names)
<span class="comment"># Concatenate into chunks of exactly context_length</span>
print(<span class="string">"Grouping into chunks..."</span>)
<span class="keyword">def</span> <span class="function">group_texts</span>(examples):
concatenated = {k: sum(examples[k], []) <span class="keyword">for</span> k <span class="keyword">in</span> examples.keys()}
total_length = len(concatenated[list(examples.keys())[0]])
<span class="comment"># Drop remainder</span>
total_length = (total_length // context_length) * context_length
result = {}
<span class="keyword">for</span> k, t <span class="keyword">in</span> concatenated.items():
result[k] = [t[i:i+context_length] <span class="keyword">for</span> i <span class="keyword">in</span> range(<span class="number">0</span>, total_length, context_length)]
result[<span class="string">"labels"</span>] = result[<span class="string">"input_ids"</span>].copy()
<span class="keyword">return</span> result
lm_dataset = tokenized.map(group_texts, batched=<span class="keyword">True</span>, num_proc=num_proc)
<span class="comment"># Split train/val</span>
lm_dataset = lm_dataset.train_test_split(test_size=<span class="number">0.001</span>, seed=<span class="variable">SEED</span>)
print(<span class="string">f"Train samples: {len(lm_dataset['train'])}"</span>)
print(<span class="string">f"Validation samples: {len(lm_dataset['test'])}"</span>)
lm_dataset.save_to_disk(output_dir)
print(<span class="string">f"Saved to {output_dir}"</span>)
<span class="keyword">return</span> lm_dataset
<span class="keyword">if</span> __name__ == <span class="string">"__main__"</span>:
prepare_pretraining_data()
</pre>
</div>
<!-- SECTION 4: TOKENIZER -->
<h2 id="tokenizer">4. Train Custom Tokenizer</h2>
<div class="note">
A custom tokenizer trained on your domain data will be more efficient than using GPT-2's. We use <strong>Byte-Pair Encoding (BPE)</strong> with the <code>tokenizers</code> library.
</div>
<div class="step">
<span class="step-number">4</span>
<strong>Train tokenizer on your corpus:</strong>
</div>
<div class="code-block">
<span class="lang">python</span>
<pre><span class="comment"># scripts/train_tokenizer.py</span>
<span class="keyword">import</span> os
<span class="keyword">from</span> datasets <span class="keyword">import</span> load_dataset
<span class="keyword">from</span> tokenizers <span class="keyword">import</span> Tokenizer, models, pre_tokenizers, trainers, processors
<span class="keyword">from</span> transformers <span class="keyword">import</span> PreTrainedTokenizerFast
<span class="variable">VOCAB_SIZE</span> = <span class="number">32000</span>
<span class="variable">CONTEXT_LENGTH</span> = <span class="number">512</span>
<span class="variable">OUTPUT_DIR</span> = <span class="string">"./tokenizer"</span>
<span class="keyword">def</span> <span class="function">train_tokenizer</span>():
os.makedirs(<span class="variable">OUTPUT_DIR</span>, exist_ok=<span class="keyword">True</span>)
print(<span class="string">"Loading training corpus for tokenizer..."</span>)
<span class="comment"># Use a sample of web text for training</span>
ds = load_dataset(<span class="string">"openwebtext"</span>, streaming=<span class="keyword">True</span>, split=<span class="string">"train"</span>)
<span class="comment"># Collect ~100MB of text for tokenizer training</span>
texts = []
total_chars = <span class="number">0</span>
target_chars = <span class="number">100_000_000</span> <span class="comment"># 100MB</span>
<span class="keyword">for</span> example <span class="keyword">in</span> ds:
text = example[<span class="string">"text"</span>]
texts.append(text)
total_chars += len(text)
<span class="keyword">if</span> total_chars >= target_chars:
<span class="keyword">break</span>
<span class="keyword">if</span> len(texts) % <span class="number">10000</span> == <span class="number">0</span>:
print(<span class="string">f"Collected {len(texts)} texts, {total_chars/1e6:.1f}MB..."</span>)
print(<span class="string">f"Training tokenizer on {len(texts)} documents ({total_chars/1e6:.1f}MB)..."</span>)
<span class="comment"># Initialize BPE tokenizer</span>
tokenizer = Tokenizer(models.BPE())
<span class="comment"># Use ByteLevel pre-tokenizer (like GPT-2) - handles all Unicode</span>
tokenizer.pre_tokenizer = pre_tokenizers.ByteLevel(add_prefix_space=<span class="keyword">False</span>)
<span class="comment"># Trainer with special tokens</span>
trainer = trainers.BpeTrainer(
vocab_size=<span class="variable">VOCAB_SIZE</span>,
special_tokens=[
<span class="string">"<|endoftext|>"</span>, <span class="comment"># EOS / PAD</span>
<span class="string">"<|bos|>"</span>, <span class="comment"># Beginning of sequence</span>
<span class="string">"<|unk|>"</span>, <span class="comment"># Unknown token</span>
<span class="string">"<|pad|>"</span>, <span class="comment"># Padding</span>
<span class="string">"<|mask|>"</span>, <span class="comment"># Mask (for potential MLM use)</span>
],
min_frequency=<span class="number">2</span>,
show_progress=<span class="keyword">True</span>,
)
<span class="comment"># Train</span>
tokenizer.train_from_iterator(texts, trainer=trainer, length=len(texts))
<span class="comment"># Add post-processor for template processing</span>
tokenizer.post_processor = processors.ByteLevel(trim_offsets=<span class="keyword">False</span>)
<span class="comment"># Enable padding and truncation</span>
tokenizer.enable_padding(
pad_id=tokenizer.token_to_id(<span class="string">"<|pad|>"</span>),
pad_token=<span class="string">"<|pad|>"</span>,
length=<span class="variable">CONTEXT_LENGTH</span>
)
tokenizer.enable_truncation(max_length=<span class="variable">CONTEXT_LENGTH</span>)
<span class="comment"># Save raw tokenizer</span>
tokenizer_path = os.path.join(<span class="variable">OUTPUT_DIR</span>, <span class="string">"tokenizer.json"</span>)
tokenizer.save(tokenizer_path)
<span class="comment"># Wrap as HuggingFace tokenizer</span>
hf_tokenizer = PreTrainedTokenizerFast(
tokenizer_object=tokenizer,
bos_token=<span class="string">"<|bos|>"</span>,
eos_token=<span class="string">"<|endoftext|>"</span>,
unk_token=<span class="string">"<|unk|>"</span>,
pad_token=<span class="string">"<|pad|>"</span>,
mask_token=<span class="string">"<|mask|>"</span>,
)
<span class="comment"># Save in HF format</span>
hf_tokenizer.save_pretrained(<span class="variable">OUTPUT_DIR</span>)
print(<span class="string">f"\n✅ Tokenizer saved to {OUTPUT_DIR}"</span>)
print(<span class="string">f"Vocab size: {len(hf_tokenizer)}"</span>)
print(<span class="string">f"Model max length: {hf_tokenizer.model_max_length}"</span>)
<span class="comment"># Test</span>
test_text = <span class="string">"Hello world! This is a test of the custom tokenizer."</span>
encoded = hf_tokenizer.encode(test_text)
decoded = hf_tokenizer.decode(encoded)
print(<span class="string">f"\nTest encode/decode:"</span>)
print(<span class="string">f" Original: {test_text}"</span>)
print(<span class="string">f" Decoded: {decoded}"</span>)
print(<span class="string">f" Tokens: {len(encoded)}"</span>)
<span class="keyword">if</span> __name__ == <span class="string">"__main__"</span>:
train_tokenizer()
</pre>
</div>
<div class="note success">
After training, your tokenizer will be saved in <code>./tokenizer/</code> with vocab size 32,000. This is more efficient than GPT-2's 50,257 vocab for general English text.
</div>
<!-- SECTION 5: MODEL CONFIG -->
<h2 id="model">5. Model Configuration</h2>
<div class="step">
<span class="step-number">5</span>
<strong>Define model architecture (~100M parameters):</strong>
</div>
<div class="code-block">
<span class="lang">python</span>
<pre><span class="comment"># scripts/create_model.py</span>
<span class="keyword">from</span> transformers <span class="keyword">import</span> GPT2Config, GPT2LMHeadModel, AutoTokenizer
<span class="keyword">import</span> os
<span class="variable">OUTPUT_DIR</span> = <span class="string">"./models/llm-100m"</span>
<span class="keyword">def</span> <span class="function">create_model_config</span>():
<span class="string">"""
Create a ~100M parameter GPT-style model.
Architecture: 12 layers, 768 hidden, 12 heads, 3072 FFN
"""</span>
config = GPT2Config(
<span class="comment"># Core architecture</span>
vocab_size=<span class="number">32000</span>, <span class="comment"># Match your tokenizer</span>
n_positions=<span class="number">512</span>, <span class="comment"># Max context length</span>
n_embd=<span class="number">768</span>, <span class="comment"># Hidden dimension</span>
n_layer=<span class="number">12</span>, <span class="comment"># Transformer layers</span>
n_head=<span class="number">12</span>, <span class="comment"># Attention heads</span>
<span class="comment"># FFN intermediate size (4× hidden = 3072)</span>
n_inner=<span class="number">3072</span>,
<span class="comment"># Activation and normalization</span>
activation_function=<span class="string">"gelu_new"</span>,
layer_norm_eps=<span class="number">1e-5</span>,
<span class="comment"># Dropout (0.1 for training, 0 for inference)</span>
resid_pdrop=<span class="number">0.1</span>,
embd_pdrop=<span class="number">0.1</span>,
attn_pdrop=<span class="number">0.1</span>,
<span class="comment"># Initialization</span>
initializer_range=<span class="number">0.02</span>,
<span class="comment"># Tie embeddings to output (saves params, often better)</span>
tie_word_embeddings=<span class="keyword">True</span>,
<span class="comment"># For gradient checkpointing compatibility</span>
use_cache=<span class="keyword">False</span>,
)
<span class="keyword">return</span> config
<span class="keyword">def</span> <span class="function">create_model</span>():
os.makedirs(<span class="variable">OUTPUT_DIR</span>, exist_ok=<span class="keyword">True</span>)
<span class="comment"># Load tokenizer</span>
tokenizer = AutoTokenizer.from_pretrained(<span class="string">"./tokenizer"</span>)
<span class="comment"># Create config</span>
config = create_model_config()
<span class="comment"># Initialize model from config</span>
model = GPT2LMHeadModel(config)
<span class="comment"># Resize embeddings to match tokenizer (in case of mismatch)</span>
model.resize_token_embeddings(len(tokenizer))
<span class="comment"># Count parameters</span>
total_params = sum(p.numel() <span class="keyword">for</span> p <span class="keyword">in</span> model.parameters())
trainable_params = sum(p.numel() <span class="keyword">for</span> p <span class="keyword">in</span> model.parameters() <span class="keyword">if</span> p.requires_grad)
print(<span class="string">f"✅ Model created!"</span>)
print(<span class="string">f" Total parameters: {total_params:,} ({total_params/1e6:.1f}M)"</span>)
print(<span class="string">f" Trainable: {trainable_params:,}"</span>)
print(<span class="string">f" Layers: {config.n_layer}"</span>)
print(<span class="string">f" Hidden: {config.n_embd}"</span>)
print(<span class="string">f" Heads: {config.n_head}"</span>)
print(<span class="string">f" Context: {config.n_positions}"</span>)
print(<span class="string">f" Vocab: {config.vocab_size}"</span>)
<span class="comment"># Save</span>
model.save_pretrained(<span class="variable">OUTPUT_DIR</span>)
tokenizer.save_pretrained(<span class="variable">OUTPUT_DIR</span>)
print(<span class="string">f"\n💾 Saved to {OUTPUT_DIR}"</span>)
<span class="keyword">return</span> model, tokenizer
<span class="keyword">if</span> __name__ == <span class="string">"__main__"</span>:
create_model()
</pre>
</div>
<div class="note">
<strong>Expected output:</strong> ~100M parameters. With tied embeddings, this is efficient for a 12-layer model.
</div>
<!-- SECTION 6: TRAINING -->
<h2 id="training">6. Training Script</h2>
<div class="step">
<span class="step-number">6</span>
<strong>Main pretraining script with optimizations:</strong>
</div>
<div class="code-block">
<span class="lang">python</span>
<pre><span class="comment"># scripts/pretrain.py</span>
<span class="keyword">import</span> os
<span class="keyword">import</span> torch
<span class="keyword">import</span> wandb
<span class="keyword">from</span> transformers <span class="keyword">import</span> (
GPT2LMHeadModel,
AutoTokenizer,
TrainingArguments,
Trainer,
DataCollatorForLanguageModeling,
get_cosine_schedule_with_warmup,
)
<span class="keyword">from</span> datasets <span class="keyword">import</span> load_from_disk
<span class="keyword">from</span> accelerate <span class="keyword">import</span> Accelerator
<span class="comment"># ==================== CONFIGURATION ====================</span>
<span class="variable">CONFIG</span> = {
<span class="string">"model_path"</span>: <span class="string">"./models/llm-100m"</span>,
<span class="string">"data_path"</span>: <span class="string">"./data/pretrain"</span>,
<span class="string">"output_dir"</span>: <span class="string">"./models/llm-100m-trained"</span>,
<span class="comment"># Training hyperparameters</span>
<span class="string">"batch_size"</span>: <span class="number">16</span>, <span class="comment"># Per device (adjust based on GPU memory)</span>
<span class="string">"gradient_accumulation"</span>: <span class="number">4</span>, <span class="comment"># Effective batch = 16 × 4 = 64</span>
<span class="string">"learning_rate"</span>: <span class="number">5e-4</span>,
<span class="string">"weight_decay"</span>: <span class="number">0.01</span>,
<span class="string">"warmup_steps"</span>: <span class="number">1000</span>,
<span class="string">"max_steps"</span>: <span class="number">50_000</span>, <span class="comment"># ~3-4 epochs on 5B tokens</span>
<span class="string">"max_grad_norm"</span>: <span class="number">1.0</span>,
<span class="comment"># Optimization</span>
<span class="string">"fp16"</span>: <span class="keyword">False</span>, <span class="comment"># Use bf16 if available (RTX 4090, A100)</span>
<span class="string">"bf16"</span>: <span class="keyword">True</span>, <span class="comment"># Better than fp16 on Ampere+ GPUs</span>
<span class="string">"gradient_checkpointing"</span>: <span class="keyword">True</span>, <span class="comment"># Trade compute for memory</span>
<span class="comment"># Logging & Checkpointing</span>
<span class="string">"logging_steps"</span>: <span class="number">100</span>,
<span class="string">"eval_steps"</span>: <span class="number">1000</span>,
<span class="string">"save_steps"</span>: <span class="number">5000</span>,
<span class="string">"save_total_limit"</span>: <span class="number">3</span>,
<span class="comment"># Other</span>
<span class="string">"seed"</span>: <span class="number">42</span>,
<span class="string">"dataloader_num_workers"</span>: <span class="number">4</span>,
}
<span class="comment"># ==================== TRAINING SETUP ====================</span>
<span class="keyword">def</span> <span class="function">setup_training</span>():
<span class="comment"># Initialize wandb (optional)</span>
wandb.init(
project=<span class="string">"llm-pretraining"</span>,
name=<span class="string">"llm-100m-general"</span>,
config=<span class="variable">CONFIG</span>,
)
<span class="comment"># Set seed</span>
torch.manual_seed(<span class="variable">CONFIG</span>[<span class="string">"seed"</span>])
<span class="comment"># Load tokenizer and model</span>
print(<span class="string">"Loading model and tokenizer..."</span>)
tokenizer = AutoTokenizer.from_pretrained(<span class="variable">CONFIG</span>[<span class="string">"model_path"</span>])
model = GPT2LMHeadModel.from_pretrained(<span class="variable">CONFIG</span>[<span class="string">"model_path"</span>])
<span class="comment"># Enable gradient checkpointing for memory savings</span>
<span class="keyword">if</span> <span class="variable">CONFIG</span>[<span class="string">"gradient_checkpointing"</span>]:
model.gradient_checkpointing_enable()
model.config.use_cache = <span class="keyword">False</span> <span class="comment"># Required when using gradient checkpointing</span>
<span class="comment"># Load dataset</span>
print(<span class="string">"Loading dataset..."</span>)
<span class="keyword">if</span> os.path.exists(<span class="variable">CONFIG</span>[<span class="string">"data_path"</span>]):
dataset = load_from_disk(<span class="variable">CONFIG</span>[<span class="string">"data_path"</span>])
<span class="keyword">else</span>:
print(<span class="string">"Dataset not found! Run prepare_data.py first."</span>)
<span class="keyword">return</span>
<span class="comment"># Data collator for causal LM (creates labels by shifting inputs)</span>
data_collator = DataCollatorForLanguageModeling(
tokenizer=tokenizer,
mlm=<span class="keyword">False</span>, <span class="comment"># Causal LM, not masked</span>
)
<span class="comment"># Training arguments</span>
training_args = TrainingArguments(
output_dir=<span class="variable">CONFIG</span>[<span class="string">"output_dir"</span>],
overwrite_output_dir=<span class="keyword">True</span>,
<span class="comment"># Batch sizes</span>
per_device_train_batch_size=<span class="variable">CONFIG</span>[<span class="string">"batch_size"</span>],
per_device_eval_batch_size=<span class="variable">CONFIG</span>[<span class="string">"batch_size"</span>],
gradient_accumulation_steps=<span class="variable">CONFIG</span>[<span class="string">"gradient_accumulation"</span>],
<span class="comment"># Learning rate</span>
learning_rate=<span class="variable">CONFIG</span>[<span class="string">"learning_rate"</span>],
weight_decay=<span class="variable">CONFIG</span>[<span class="string">"weight_decay"</span>],
max_grad_norm=<span class="variable">CONFIG</span>[<span class="string">"max_grad_norm"</span>],
warmup_steps=<span class="variable">CONFIG</span>[<span class="string">"warmup_steps"</span>],
<span class="comment"># Schedule</span>
lr_scheduler_type=<span class="string">"cosine"</span>,
max_steps=<span class="variable">CONFIG</span>[<span class="string">"max_steps"</span>],
<span class="comment"># Precision</span>
fp16=<span class="variable">CONFIG</span>[<span class="string">"fp16"</span>],
bf16=<span class="variable">CONFIG</span>[<span class="string">"bf16"</span>],
<span class="comment"># Evaluation</span>
evaluation_strategy=<span class="string">"steps"</span>,
eval_steps=<span class="variable">CONFIG</span>[<span class="string">"eval_steps"</span>],
<span class="comment"># Logging</span>
logging_strategy=<span class="string">"steps"</span>,
logging_steps=<span class="variable">CONFIG</span>[<span class="string">"logging_steps"</span>],
report_to=[<span class="string">"wandb"</span>],
<span class="comment"># Checkpointing</span>
save_strategy=<span class="string">"steps"</span>,
save_steps=<span class="variable">CONFIG</span>[<span class="string">"save_steps"</span>],
save_total_limit=<span class="variable">CONFIG</span>[<span class="string">"save_total_limit"</span>],
<span class="comment"># Performance</span>
dataloader_num_workers=<span class="variable">CONFIG</span>[<span class="string">"dataloader_num_workers"</span>],
dataloader_pin_memory=<span class="keyword">True</span>,
remove_unused_columns=<span class="keyword">False</span>,
<span class="comment"># Hub</span>
push_to_hub=<span class="keyword">False</span>, <span class="comment"># Set True to upload checkpoints</span>
hub_model_id=<span class="keyword">None</span>,
seed=<span class="variable">CONFIG</span>[<span class="string">"seed"</span>],
)
<span class="comment"># Initialize Trainer</span>
trainer = Trainer(
model=model,
args=training_args,
train_dataset=dataset[<span class="string">"train"</span>],
eval_dataset=dataset[<span class="string">"test"</span>],
data_collator=data_collator,
)
<span class="comment"># Train</span>
print(<span class="string">"\n🚀 Starting training...\n"</span>)
trainer.train()
<span class="comment"># Save final model</span>
trainer.save_model(<span class="variable">CONFIG</span>[<span class="string">"output_dir"</span>])
tokenizer.save_pretrained(<span class="variable">CONFIG</span>[<span class="string">"output_dir"</span>])
<span class="comment"># Final evaluation</span>
eval_results = trainer.evaluate()
print(<span class="string">f"\n📊 Final eval loss: {eval_results['eval_loss']:.4f}"</span>)
print(<span class="string">f"📊 Final perplexity: {torch.exp(torch.tensor(eval_results['eval_loss'])).item():.2f}"</span>)
wandb.finish()
<span class="keyword">if</span> __name__ == <span class="string">"__main__"</span>:
setup_training()
</pre>
</div>
<div class="note">
<strong>Run training:</strong> <code class="command">python scripts/pretrain.py</code>
</div>
<!-- SECTION 7: OPTIMIZATION -->
<h2 id="optimization">7. Memory Optimizations for Single GPU</h2>
<div class="note warning">
If you run out of memory, use these techniques in order of impact:
</div>
<table>
<tr><th>Technique</th><th>Memory Saved</th><th>How to Enable</th></tr>
<tr><td>Gradient Checkpointing</td><td>~60%</td><td><code>model.gradient_checkpointing_enable()</code></td></tr>
<tr><td>BF16 / FP16</td><td>~50%</td><td><code>bf16=True</code> in TrainingArguments</td></tr>
<tr><td>Gradient Accumulation</td><td>Scales effective batch</td><td><code>gradient_accumulation_steps=4+</code></td></tr>
<tr><td>Smaller Batch Size</td><td>Linear</td><td><code>per_device_train_batch_size=4</code></td></tr>
<tr><td>DeepSpeed ZeRO-2</td><td>~20%</td><td>Use DeepSpeed config (see below)</td></tr>
<tr><td>8-bit AdamW</td><td>~15%</td><td><code>pip install bitsandbytes</code>, use <code>bnb.optim.AdamW8bit</code></td></tr>
</table>
<h3>DeepSpeed Config (for aggressive memory optimization)</h3>
<div class="code-block">
<span class="lang">json</span>
<pre><span class="comment"># ds_config.json</span>
{
<span class="string">"bf16"</span>: { <span class="string">"enabled"</span>: <span class="keyword">true</span> },
<span class="string">"zero_optimization"</span>: {
<span class="string">"stage"</span>: <span class="number">2</span>,
<span class="string">"offload_optimizer"</span>: {
<span class="string">"device"</span>: <span class="string">"cpu"</span>,
<span class="string">"pin_memory"</span>: <span class="keyword">true</span>
},
<span class="string">"allgather_partitions"</span>: <span class="keyword">true</span>,
<span class="string">"allgather_bucket_size"</span>: <span class="number">2e8</span>,
<span class="string">"overlap_comm"</span>: <span class="keyword">true</span>,
<span class="string">"reduce_scatter"</span>: <span class="keyword">true</span>,
<span class="string">"reduce_bucket_size"</span>: <span class="number">2e8</span>
},
<span class="string">"train_batch_size"</span>: <span class="string">"auto"</span>,
<span class="string">"train_micro_batch_size_per_gpu"</span>: <span class="string">"auto"</span>,
<span class="string">"gradient_accumulation_steps"</span>: <span class="string">"auto"</span>,
<span class="string">"optimizer"</span>: {
<span class="string">"type"</span>: <span class="string">"AdamW"</span>,
<span class="string">"params"</span>: {
<span class="string">"lr"</span>: <span class="number">5e-4</span>,
<span class="string">"betas"</span>: [<span class="number">0.9</span>, <span class="number">0.999</span>],
<span class="string">"eps"</span>: <span class="number">1e-8</span>,
<span class="string">"weight_decay"</span>: <span class="number">0.01</span>
}
},
<span class="string">"scheduler"</span>: {
<span class="string">"type"</span>: <span class="string">"WarmupLR"</span>,
<span class="string">"params"</span>: { <span class="string">"warmup_min_lr"</span>: <span class="number">0</span>, <span class="string">"warmup_max_lr"</span>: <span class="number">5e-4</span>, <span class="string">"warmup_num_steps"</span>: <span class="number">1000</span> }
}
}
</pre>
</div>
<div class="note">
Run with DeepSpeed: <code class="command">deepspeed scripts/pretrain.py --deepspeed ds_config.json</code>
</div>
<!-- SECTION 8: EVALUATION -->
<h2 id="evaluation">8. Evaluation & Benchmarking</h2>
<div class="step">
<span class="step-number">7</span>
<strong>Evaluate your trained model:</strong>
</div>
<div class="code-block">
<span class="lang">python</span>
<pre><span class="comment"># scripts/evaluate.py</span>
<span class="keyword">import</span> torch
<span class="keyword">from</span> transformers <span class="keyword">import</span> GPT2LMHeadModel, AutoTokenizer
<span class="keyword">from</span> datasets <span class="keyword">import</span> load_dataset
<span class="keyword">import</span> math
<span class="keyword">def</span> <span class="function">evaluate_perplexity</span>(model_path=<span class="string">"./models/llm-100m-trained"</span>):
<span class="string">"""Calculate perplexity on validation set."""</span>
tokenizer = AutoTokenizer.from_pretrained(model_path)
model = GPT2LMHeadModel.from_pretrained(model_path)
model.eval()
<span class="comment"># Load a small validation set</span>
ds = load_dataset(<span class="string">"wikitext"</span>, <span class="string">"wikitext-2-raw-v1"</span>, split=<span class="string">"validation"</span>)
encodings = tokenizer(<span class="string">"\n\n"</span>.join(ds[<span class="string">"text"</span>]), return_tensors=<span class="string">"pt"</span>)
max_length = <span class="number">512</span>
stride = <span class="number">512</span>
seq_len = encodings.input_ids.size(<span class="number">1</span>)
nlls = []
prev_end_loc = <span class="number">0</span>
<span class="keyword">for</span> begin_loc <span class="keyword">in</span> range(<span class="number">0</span>, seq_len, stride):
end_loc = min(begin_loc + max_length, seq_len)
trg_len = end_loc - prev_end_loc
input_ids = encodings.input_ids[:, begin_loc:end_loc]
target_ids = input_ids.clone()
target_ids[:, :-trg_len] = -<span class="number">100</span>
<span class="keyword">with</span> torch.no_grad():
outputs = model(input_ids, labels=target_ids)
neg_log_likelihood = outputs.loss * trg_len
nlls.append(neg_log_likelihood)
prev_end_loc = end_loc
<span class="keyword">if</span> end_loc == seq_len:
<span class="keyword">break</span>
ppl = torch.exp(torch.stack(nlls).sum() / end_loc)
print(<span class="string">f"Perplexity: {ppl.item():.2f}"</span>)
<span class="keyword">return</span> ppl.item()
<span class="keyword">def</span> <span class="function">generate_samples</span>(model_path=<span class="string">"./models/llm-100m-trained"</span>, prompts=<span class="keyword">None</span>):
<span class="string">"""Generate text samples to qualitatively evaluate."""</span>
<span class="keyword">if</span> prompts <span class="keyword">is</span> <span class="keyword">None</span>:
prompts = [
<span class="string">"The future of artificial intelligence is"</span>,
<span class="string">"Once upon a time in a distant galaxy"</span>,
<span class="string">"The key to happiness lies in"</span>,
<span class="string">"In the field of machine learning,"</span>,
<span class="string">"The quick brown fox"</span>,
]
tokenizer = AutoTokenizer.from_pretrained(model_path)
model = GPT2LMHeadModel.from_pretrained(model_path)
model.eval()
device = <span class="string">"cuda"</span> <span class="keyword">if</span> torch.cuda.is_available() <span class="keyword">else</span> <span class="string">"cpu"</span>
model = model.to(device)
print(<span class="string">"="</span>*<span class="number">60</span>)
<span class="keyword">for</span> prompt <span class="keyword">in</span> prompts:
inputs = tokenizer(prompt, return_tensors=<span class="string">"pt"</span>).to(device)
<span class="keyword">with</span> torch.no_grad():
outputs = model.generate(
**inputs,
max_new_tokens=<span class="number">100</span>,
do_sample=<span class="keyword">True</span>,
temperature=<span class="number">0.8</span>,
top_k=<span class="number">50</span>,
top_p=<span class="number">0.95</span>,
num_return_sequences=<span class="number">1</span>,
pad_token_id=tokenizer.eos_token_id,
)
generated = tokenizer.decode(outputs[<span class="number">0</span>], skip_special_tokens=<span class="keyword">True</span>)
print(<span class="string">f"\n📝 Prompt: {prompt}"</span>)
print(<span class="string">f"🤖 Output: {generated}"</span>)
print(<span class="string">"-"</span>*<span class="number">60</span>)
<span class="keyword">if</span> __name__ == <span class="string">"__main__"</span>:
print(<span class="string">"Evaluating perplexity..."</span>)
evaluate_perplexity()
print(<span class="string">"\nGenerating samples..."</span>)
generate_samples()
</pre>
</div>
<!-- SECTION 9: INFERENCE -->
<h2 id="inference">9. Inference & Text Generation</h2>
<div class="step">
<span class="step-number">8</span>
<strong>Simple inference script:</strong>
</div>
<div class="code-block">
<span class="lang">python</span>
<pre><span class="comment"># scripts/generate.py</span>
<span class="keyword">import</span> torch
<span class="keyword">from</span> transformers <span class="keyword">import</span> pipeline
<span class="keyword">def</span> <span class="function">create_generator</span>(model_path=<span class="string">"./models/llm-100m-trained"</span>):
<span class="string">"""Create a text generation pipeline."""</span>
device = <span class="string">"cuda"</span> <span class="keyword">if</span> torch.cuda.is_available() <span class="keyword">else</span> <span class="string">"cpu"</span>
generator = pipeline(
<span class="string">"text-generation"</span>,
model=model_path,
tokenizer=model_path,
device=device,
torch_dtype=torch.bfloat16 <span class="keyword">if</span> device == <span class="string">"cuda"</span> <span class="keyword">else</span> torch.float32,
)
<span class="keyword">return</span> generator
<span class="keyword">def</span> <span class="function">chat_loop</span>():
<span class="string">"""Interactive text generation."""</span>
print(<span class="string">"Loading model..."</span>)
generator = create_generator()
print(<span class="string">"\n🤖 LLM Ready! Type your prompt (or 'quit' to exit)\n"</span>)
<span class="keyword">while</span> <span class="keyword">True</span>:
prompt = input(<span class="string">"You: "</span>).strip()
<span class="keyword">if</span> prompt.lower() <span class="keyword">in</span> [<span class="string">"quit"</span>, <span class="string">"exit"</span>, <span class="string">"q"</span>]:
<span class="keyword">break</span>
<span class="keyword">if</span> <span class="keyword">not</span> prompt:
<span class="keyword">continue</span>
<span class="comment"># Generate</span>
outputs = generator(
prompt,
max_new_tokens=<span class="number">150</span>,
do_sample=<span class="keyword">True</span>,
temperature=<span class="number">0.8</span>,
top_p=<span class="number">0.92</span>,
top_k=<span class="number">50</span>,
repetition_penalty=<span class="number">1.1</span>,
pad_token_id=generator.tokenizer.eos_token_id,
)
generated_text = outputs[<span class="number">0</span>][<span class="string">"generated_text"</span>]
response = generated_text[len(prompt):].strip()
print(<span class="string">f"\n🤖 Bot: {response}\n"</span>)
<span class="keyword">if</span> __name__ == <span class="string">"__main__"</span>:
chat_loop()
</pre>
</div>
<!-- SECTION 10: ALIGNMENT -->
<h2 id="alignment">10. Optional: SFT / DPO Alignment</h2>
<div class="note success">
After pretraining, align your model with instruction-following data using <strong>TRL</strong>.
</div>
<h3>Supervised Fine-Tuning (SFT)</h3>
<div class="code-block">
<span class="lang">python</span>
<pre><span class="comment"># scripts/sft_train.py</span>
<span class="keyword">from</span> transformers <span class="keyword">import</span> AutoModelForCausalLM, AutoTokenizer, TrainingArguments
<span class="keyword">from</span> trl <span class="keyword">import</span> SFTTrainer, DataCollatorForCompletionOnlyLM
<span class="keyword">from</span> datasets <span class="keyword">import</span> load_dataset
<span class="comment"># Load pretrained model</span>
model = AutoModelForCausalLM.from_pretrained(<span class="string">"./models/llm-100m-trained"</span>)
tokenizer = AutoTokenizer.from_pretrained(<span class="string">"./models/llm-100m-trained"</span>)
<span class="comment"># Load instruction dataset (e.g., Alpaca, Dolly, OpenAssistant)</span>
dataset = load_dataset(<span class="string">"tatsu-lab/alpaca"</span>, split=<span class="string">"train"</span>)
<span class="comment"># Format: "### Instruction:\n{instruction}\n\n### Response:\n{output}"</span>
<span class="keyword">def</span> <span class="function">format_prompt</span>(example):
<span class="keyword">if</span> example[<span class="string">"input"</span>]:
prompt = <span class="string">f"### Instruction:\n{example['instruction']}\n\n### Input:\n{example['input']}\n\n### Response:\n{example['output']}"</span>
<span class="keyword">else</span>:
prompt = <span class="string">f"### Instruction:\n{example['instruction']}\n\n### Response:\n{example['output']}"</span>
<span class="keyword">return</span> {<span class="string">"text"</span>: prompt}
dataset = dataset.map(format_prompt)
<span class="comment"># Train with SFTTrainer</span>
training_args = TrainingArguments(
output_dir=<span class="string">"./models/llm-100m-sft"</span>,
num_train_epochs=<span class="number">3</span>,
per_device_train_batch_size=<span class="number">8</span>,
learning_rate=<span class="number">2e-5</span>,
fp16=<span class="keyword">True</span>,
logging_steps=<span class="number">10</span>,
save_strategy=<span class="string">"epoch"</span>,
)
trainer = SFTTrainer(
model=model,
tokenizer=tokenizer,
train_dataset=dataset,
args=training_args,
dataset_text_field=<span class="string">"text"</span>,
max_seq_length=<span class="number">512</span>,
)
trainer.train()
trainer.save_model(<span class="string">"./models/llm-100m-sft"</span>)
</pre>
</div>
<h3>Direct Preference Optimization (DPO)</h3>
<div class="code-block">
<span class="lang">python</span>
<pre><span class="comment"># scripts/dpo_train.py</span>
<span class="keyword">from</span> transformers <span class="keyword">import</span> AutoModelForCausalLM, AutoTokenizer
<span class="keyword">from</span> trl <span class="keyword">import</span> DPOTrainer, DPOConfig
<span class="keyword">from</span> datasets <span class="keyword">import</span> load_dataset
<span class="comment"># Load SFT model as base</span>
model = AutoModelForCausalLM.from_pretrained(<span class="string">"./models/llm-100m-sft"</span>)
tokenizer = AutoTokenizer.from_pretrained(<span class="string">"./models/llm-100m-sft"</span>)
<span class="comment"># Load preference dataset (e.g., Anthropic HH, SHP)</span>
dataset = load_dataset(<span class="string">"trl-lib/ultrafeedback_binarized"</span>, split=<span class="string">"train"</span>)
<span class="comment"># DPO training</span>
training_args = DPOConfig(
output_dir=<span class="string">"./models/llm-100m-dpo"</span>,
num_train_epochs=<span class="number">1</span>,
per_device_train_batch_size=<span class="number">4</span>,
learning_rate=<span class="number">1e-6</span>,
beta=<span class="number">0.1</span>, <span class="comment"># DPO temperature parameter</span>
logging_steps=<span class="number">10</span>,
)
trainer = DPOTrainer(
model=model,
args=training_args,
train_dataset=dataset,
tokenizer=tokenizer,
)
trainer.train()
trainer.save_model(<span class="string">"./models/llm-100m-dpo"</span>)
</pre>
</div>
<!-- QUICK START -->
<h2>🚀 Quick Start Commands</h2>
<div class="code-block">
<span class="lang">bash</span>
<pre><span class="comment"># 1. Setup</span>
mkdir llm-project && cd llm-project
python -m venv venv
source venv/bin/activate
pip install -r requirements.txt
<span class="comment"># 2. Train tokenizer</span>
python scripts/train_tokenizer.py
<span class="comment"># 3. Prepare data</span>
python scripts/prepare_data.py
<span class="comment"># 4. Create model</span>
python scripts/create_model.py
<span class="comment"># 5. Pretrain!</span>
python scripts/pretrain.py
<span class="comment"># 6. Evaluate</span>
python scripts/evaluate.py
<span class="comment"># 7. Chat with your model</span>
python scripts/generate.py
<span class="comment"># 8. (Optional) Align with SFT</span>
python scripts/sft_train.py
</pre>
</div>
<div class="note success">
<strong>Expected Results:</strong> After ~50K steps on 5B tokens, you should see perplexity around <strong>15-25</strong> on WikiText-2. The model will generate coherent sentences and understand basic context.
</div>
<div class="footer">
<p>Built with 🤗 Hugging Face ecosystem | Transformers + Datasets + Tokenizers + TRL + Accelerate</p>
<p>For issues or questions, check the <a href="https://huggingface.co/learn/llm-course">LLM Course</a> and <a href="https://huggingface.co/docs/transformers">Transformers Docs</a></p>
</div>
</div>
</body>
</html>
<script data-huggingchat-badge="script">
(function () {
try {
if (document.querySelector("[data-huggingchat-badge]:not(script)")) return;
var host = document.createElement("div");
host.setAttribute("data-huggingchat-badge", "");
var styles = [["all","initial"],["position","fixed"],["right","12px"],["bottom","12px"],["z-index","2147483647"],["display","block"],["visibility","visible"],["opacity","1"],["pointer-events","auto"],["width","auto"],["height","auto"],["margin","0"],["padding","0"],["max-width","none"],["max-height","none"],["transform","none"],["filter","none"],["clip-path","none"],["color-scheme","light"]];
for (var i = 0; i < styles.length; i++) {
host.style.setProperty(styles[i][0], styles[i][1], "important");
}
host.attachShadow({ mode: "closed" }).innerHTML = "<style>\n:host { all: initial; }\na {\n display: flex;\n align-items: center;\n gap: 5px;\n box-sizing: border-box;\n height: 26px;\n padding: 0 9px 0 7px;\n border-radius: 999px;\n border: 1px solid rgba(0, 0, 0, 0.07);\n background: rgba(255, 255, 255, 0.92);\n -webkit-backdrop-filter: saturate(180%) blur(8px);\n backdrop-filter: saturate(180%) blur(8px);\n box-shadow: 0 1px 2px rgba(0, 0, 0, 0.08), 0 6px 16px rgba(0, 0, 0, 0.06);\n color: #111827;\n font: 500 11px/1 ui-sans-serif, system-ui, -apple-system, \"Segoe UI\", sans-serif;\n letter-spacing: 0.01em;\n text-decoration: none;\n white-space: nowrap;\n opacity: 0.72;\n transition: opacity 0.15s ease, transform 0.15s ease;\n}\na:hover, a:focus-visible { opacity: 1; transform: translateY(-1px); }\nsvg { width: 14px; height: 14px; flex: none; }\n@media print { a { display: none; } }\n<\/style><a href=\"https://huggingface.co/chat\" target=\"_blank\" rel=\"noopener noreferrer nofollow\"><svg viewBox=\"0 0 32 32\" fill=\"none\" aria-hidden=\"true\"><path d=\"M16.0006 25.9992C13.8266 25.999 11.7118 25.2901 9.97686 23.9799C8.2419 22.6698 6.98127 20.8298 6.38599 18.7388C5.79071 16.6478 5.89323 14.4198 6.678 12.3923C7.46278 10.3648 8.88705 8.64837 10.735 7.50308C12.5829 6.35779 14.7538 5.84606 16.9187 6.04544C19.0837 6.24481 21.1246 7.14442 22.7323 8.60795C24.34 10.0715 25.4268 12.0192 25.8281 14.1559C26.2293 16.2926 25.9232 18.5019 24.9561 20.449C24.7703 20.8042 24.7223 21.2155 24.8211 21.604L25.4211 23.8316C25.4803 24.0518 25.4805 24.2837 25.4216 24.5039C25.3627 24.7242 25.2468 24.925 25.0856 25.0862C24.9244 25.2474 24.7235 25.3633 24.5033 25.4222C24.283 25.4811 24.0512 25.4809 23.831 25.4217L21.6034 24.8217C21.2172 24.7248 20.809 24.7729 20.4558 24.9567C19.0683 25.6467 17.5457 26.0068 16.0006 26.0068V25.9992Z\" fill=\"currentColor\"/><path d=\"M9.62598 16.0013C9.62598 15.3799 10.1294 14.8765 10.7508 14.8765C11.3721 14.8765 11.8756 15.3799 11.8756 16.0013C11.8756 17.0953 12.3102 18.1448 13.0838 18.9184C13.8574 19.692 14.9069 20.1266 16.001 20.1267C17.095 20.1267 18.1445 19.692 18.9181 18.9184C19.6918 18.1448 20.1264 17.0953 20.1264 16.0013C20.1264 15.3799 20.6299 14.8765 21.2512 14.8765C21.8725 14.8765 22.3759 15.3799 22.3759 16.0013C22.3759 17.6921 21.7046 19.3137 20.509 20.5093C19.3134 21.7049 17.6918 22.3762 16.001 22.3762C14.3102 22.3762 12.6885 21.7049 11.4929 20.5093C10.2974 19.3137 9.62598 17.6921 9.62598 16.0013Z\" fill=\"#fff\"/><\/svg><span>Made with HuggingChat<\/span><\/a>";
function attach() {
var parent = document.body || document.documentElement;
if (parent && host.parentNode !== parent) parent.appendChild(host);
}
attach();
// Artifacts that rewrite document.body wholesale would drop the badge.
if (document.body && typeof MutationObserver === "function") {
new MutationObserver(attach).observe(document.body, { childList: true });
} else {
document.addEventListener("DOMContentLoaded", attach);
}
} catch (e) {}
})();
</script>
|