Spaces:
Running
Running
File size: 31,459 Bytes
e137e27 005657d e137e27 8262fca e137e27 005657d 87a6313 e137e27 a4dc57a e137e27 e384d00 005657d e384d00 ddc7526 e384d00 e74bc72 e384d00 e74bc72 e384d00 ddc7526 e384d00 4cc0103 ddc7526 02247f7 e384d00 e74bc72 85b59b6 e384d00 ddc7526 e384d00 ddc7526 e384d00 ddc7526 e384d00 ddc7526 e384d00 85b59b6 e384d00 ddc7526 e384d00 ddc7526 e384d00 5b6d908 005657d e384d00 005657d 35a3f42 5a77be1 85b59b6 5a77be1 35a3f42 005657d e137e27 e384d00 e137e27 9a127b5 e137e27 e384d00 e137e27 43e1d29 45ddd25 43e1d29 e137e27 5d3f993 09bef6a e137e27 a4dc57a 09bef6a e137e27 b2b380b 09bef6a e137e27 fac35b0 4cc0103 09bef6a fac35b0 e137e27 db08107 09bef6a db08107 09bef6a db08107 09bef6a db08107 09bef6a db08107 09bef6a db08107 e137e27 fac35b0 4cc0103 09bef6a fac35b0 e137e27 48d8ec3 09bef6a 48d8ec3 09bef6a 48d8ec3 09bef6a 48d8ec3 e137e27 fac35b0 9a127b5 09bef6a ec2b3ce fd8de54 09bef6a fd8de54 a4dc57a 09bef6a fd8de54 09bef6a fd8de54 09bef6a fd8de54 09bef6a fd8de54 09bef6a fd8de54 0595e24 09bef6a fd8de54 0595e24 09bef6a fd8de54 0595e24 fd8de54 ec2b3ce d673af7 09bef6a fac35b0 e137e27 ea708b9 09bef6a ea708b9 a4dc57a 09bef6a ea708b9 09bef6a ea708b9 a711d2f ea708b9 e137e27 09bef6a 45ddd25 09bef6a e137e27 35a3f42 005657d e137e27 8061116 adcd5e6 8061116 adcd5e6 8061116 adcd5e6 8061116 adcd5e6 8061116 adcd5e6 8061116 adcd5e6 8061116 adcd5e6 8061116 4e6ee79 8061116 adcd5e6 8061116 4e6ee79 8061116 e384d00 8061116 adcd5e6 8061116 e384d00 8061116 3f67a06 e384d00 861154a 3f67a06 e384d00 8061116 5614f01 dbbb9f4 58a867d 8061116 dbbb9f4 5614f01 82df62a 5614f01 8061116 8580754 fac35b0 8580754 12ce41f fac35b0 2ecaabf fac35b0 2ecaabf fac35b0 b6d74c9 140edc3 8580754 fac35b0 12ce41f fac35b0 dbbb9f4 fac35b0 b6d74c9 a1001c2 140edc3 dbbb9f4 8580754 fac35b0 89cfcec fac35b0 56335b8 fac35b0 56335b8 fac35b0 89cfcec fac35b0 56335b8 fac35b0 56335b8 fac35b0 56335b8 fac35b0 7ab95df fac35b0 56335b8 fac35b0 56335b8 fac35b0 12ce41f e3b3325 f754e2b d6d69e3 f754e2b e3b3325 0e10a03 a1001c2 d6d69e3 e384d00 d6d69e3 2d4ad39 e384d00 8580754 4cc0103 0595e24 fac35b0 a4dc57a fac35b0 9a127b5 e384d00 ac7d8cf 4cc0103 e384d00 0595e24 fac35b0 9cbd894 9345dee 9aaf568 9cbd894 09bef6a e137e27 5025d3d ac7d8cf 4cc0103 0595e24 5025d3d 8061116 e384d00 3d1994e 4cc0103 3d1994e e384d00 0595e24 4cc0103 e384d00 09bef6a 5025d3d 4cc0103 5025d3d 4cc0103 5025d3d 3d1994e 5025d3d 4cc0103 5025d3d 117a05e 5025d3d 4cc0103 5025d3d e384d00 3d1994e e384d00 09bef6a 5025d3d e137e27 005657d 87a6313 e137e27 |
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 |
from fasthtml.common import *
from fasthtml.components import *
from fasthtml.components import (
D_title,
D_article,
D_front_matter,
D_contents,
D_byline,
D_bibliography,
D_appendix,
D_cite,
)
from plotly import graph_objects as go
from fh_plotly import plotly2fasthtml
import pandas as pd
import json
from rich import print
import overview
import curated
import web
import common
import results
from pybtex.database import parse_file
import data_viewer
from eval_result_figures import all_eval_res_figs
app, rt = fast_app(
debug=True,
pico=False,
hdrs=(
Meta(charset="UTF-8"),
Meta(name="viewport", content="width=device-width, initial-scale=1.0"),
Script(src="https://distill.pub/template.v2.js"),
Script(src="https://unpkg.com/htmx.org@next/dist/htmx.min.js"),
Script(src="https://cdn.plot.ly/plotly-latest.min.js"),
Link(rel="stylesheet", href="style.css"),
MarkdownJS(),
),
)
front_matter = {
"title": "TxT360",
"description": "A globally deduplicated dataset for LLM pretraining",
"published": "October 7, 2024",
"authors": [
{
"author": "Liping Tang",
"authorURL": "https://huggingface.co/Liping",
"affiliation": "MBZUAI",
"affiliationURL": "LLM360.ai",
},
{
"author": "Nikhil Ranjan",
"authorURL": "https://huggingface.co/nikhilranjan",
"affiliation": "MBZUAI",
"affiliationURL": "LLM360.ai",
},
{
"author": "Omkar Pangarkar",
"authorURL": "https://huggingface.co/omkarenator",
"affiliation": "Petuum, Inc.",
"affiliationURL": "",
},
{
"author": "Xuezhi Liang",
"authorURL": "",
"affiliation": "MBZUAI",
"affiliationURL": "",
},
{
"author": "Zhen Wang",
"authorURL": "",
"affiliation": "MBZUAI",
"affiliationURL": "",
},
{
"author": "Li An",
"authorURL": "https://huggingface.co/an1118",
"affiliation": "UCSB",
"affiliationURL": "",
},
{
"author": "Bhaskar Rao",
"authorURL": "",
"affiliation": "MBZUAI",
"affiliationURL": "",
},
{
"author": "Linghao Jin",
"authorURL": "https://sites.google.com/usc.edu/linghao-jin",
"affiliation": "USC",
"affiliationURL": "",
},
{
"author": "Huijuan Wang",
"authorURL": "https://sites.google.com/usc.edu/huijuanwang",
"affiliation": "USC",
"affiliationURL": "",
},
{
"author": "Zhoujun Cheng",
"authorURL": "https://huggingface.co/zhoujun",
"affiliation": "UCSD",
"affiliationURL": "",
},
{
"author": "Suqi Sun",
"authorURL": "https://huggingface.co/mylibrar",
"affiliation": "Petuum, Inc.",
"affiliationURL": "",
},
{
"author": "Cun Mu",
"authorURL": "https://huggingface.co/CarisMu",
"affiliation": "MBZUAI",
"affiliationURL": "",
},
{
"author": "Victor Miller",
"authorURL": "https://huggingface.co/vamiller12",
"affiliation": "Petuum, Inc.",
"affiliationURL": "",
},
{
"author": "Xuezhe Ma",
"authorURL": "https://xuezhemax.github.io/",
"affiliation": "USC",
"affiliationURL": "",
},
{
"author": "Yue Peng",
"authorURL": "https://huggingface.co/Dreamever",
"affiliation": "MBZUAI",
"affiliationURL": "",
},
{
"author": "Zhengzhong Liu",
"authorURL": "https://huggingface.co/hunterhector",
"affiliation": "Petuum, Inc. / MBZUAI ",
"affiliationURL": "",
},
{
"author": "Eric P. Xing",
"authorURL": "",
"affiliation": "MBZUAI",
"affiliationURL": "https://www.mbzuai.ac.ae/ & https://www.cs.cmu.edu/",
},
],
"katex": {"delimiters": [{"left": "$$", "right": "$$", "display": "false"}]},
}
citation_long = """
@misc{txt360data2024,
title={TxT360: A Top-Quality LLM Pre-training Dataset Requires the Perfect Blend},
author={Liping Tang, Nikhil Ranjan, Omkar Pangarkar, Xuezhi Liang, Zhen Wang, Li An, Bhaskar Rao, Linghao Jin, Huijuan Wang, Zhoujun Cheng, Suqi Sun, Cun Mu, Victor Miller, Xuezhe Ma, Yue Peng, Zhengzhong Liu, Eric P. Xing},
year={2024}
}
"""
def read_bibs():
bib_data = parse_file("bibliography.bib")
cits = []
for key in bib_data.entries.keys():
cits.append(D_cite(bibtex_key=key))
return cits
@app.get("/bibliography.bib")
def get():
return FileResponse("bibliography.bib")
@app.get("/")
def main():
from fasthtml.xtend import Script
return Div(
D_title(
H1(
"TxT360: A Top-Quality LLM Pre-training Dataset Requires the Perfect Blend",
cls="l-body",
style="text-align: center;",
),
Div(
Img(src="images/llm360_logo.png"),
id="title-plot",
cls="main-plot-container l-page",
),
),
D_byline(),
D_front_matter(
Script(
json.dumps(front_matter),
id="distill-front-matter",
type="text/json",
)
),
D_article(
D_contents(
Nav(
H3("Table of Contents"),
Div(
A(
"TxT360",
href="#section11",
)
),
Div(
Ul(
Li(
A(
"About TxT360",
href="#section11",
)
),
Li(
A(
"Why TxT360",
href="#section12",
)
),
Li(
A(
"Generalizable Approach to Data Processing",
href="#section13",
)
),
),
),
Div(
A(
"Common Crawl Data",
href="#section21",
)
),
Div(
Ul(
Li(
A(
"Common Crawl Snapshot Processing",
href="#section21",
)
),
Li(
A(
"Common Crawl Data Processing Summary",
href="#section22",
)
),
Li(
A(
"Document Preparation",
href="#section23",
)
),
Li(
A(
"Line-Level Removal",
href="#section24",
)
),
Li(
A(
"Document-Level Filtering",
href="#section25",
)
),
),
),
Div(
A(
"Curated Sources",
href="#section31",
)
),
Div(
Ul(
Li(
A(
"Curated Sources in TxT360",
href="#section31",
)
),
Li(
A(
"Filtering Steps and Definitions",
href="#section32",
)
),
Li(
A(
"Filtering Discussion on All Curated Sources",
href="#section33",
)
),
),
),
Div(
A(
"Shared Processing Steps",
href="#section41",
)
),
Div(
Ul(
Li(
A(
"Overview",
href="#section41",
)
),
Li(
A(
"Why Global Deduplication",
href="#section42",
)
),
Li(
A(
"MinHash Generation",
href="#section43",
)
),
Li(
A(
"Matching Pairs Generation",
href="#section44",
)
),
Li(
A(
"Finding Duplicate Pairs",
href="#section45",
)
),
Li(
A(
"Finding Connected Components using MapReduce",
href="#section46",
)
),
Li(
A(
"Analysis of Near-Duplicate Clusters",
href="#section47",
)
),
Li(
A(
"Personally Identifiable Information Removal",
href="#section48",
)
),
Li(
A(
"Normalization Form C",
href="#section49",
)
),
),
),
Div(
A(
"TxT360 Studies",
href="#section51",
),
),
Div(
Ul(
Li(
A(
"Overview",
href="#section51",
)
),
Li(
A(
"A Simple Data Mix Creates a Good Learning Curve",
href="#section52",
)
),
Li(
A(
"Perplexity Analysis",
href="#section53",
)
),
Li(
A(
"Topic Analysis",
href="#section55",
)
)
),
),
role="navigation",
cls="l-text figcaption",
),
),
intro(),
web.web_data(),
curated.curated(),
common.common_steps(),
results.results(),
),
D_appendix(
D_bibliography(src="bibliography.bib"),
H3("Citation"),
P("For attribution in academic contexts, please cite this work as"),
Pre(citation_long, cls="citation long"),
),
Div(*read_bibs(), style="display: none;"),
)
new_dataset_comparison1 = pd.DataFrame(
{
"Data Source": [
"CommonCrawl Snapshots",
"Papers",
"Wikipedia",
"FreeLaw",
"DM Math",
"USPTO",
"PG-19",
"HackerNews",
"Ubuntu IRC",
"EuroParl",
"StackExchange",
"Code",
],
"TxT360": [
"99",
"5 Sources",
"310+ Languages",
"Included",
"Included",
"Included",
"Included",
"Included",
"Included",
"Included",
"Included",
"**",
],
"FineWeb": [
"96",
"-",
"-",
"-",
"-",
"-",
"-",
"-",
"-",
"-",
"-",
"-",
],
"RefinedWeb": [
"90",
"-",
"-",
"-",
"-",
"-",
"-",
"-",
"-",
"-",
"-",
"-",
],
"PedPajamaV2": [
"84",
"-",
"-",
"-",
"-",
"-",
"-",
"-",
"-",
"-",
"-",
"-",
],
"C4": [
"1",
"-",
"-",
"-",
"-",
"-",
"-",
"-",
"-",
"-",
"-",
"-",
],
"Dolma": [
"24",
"1 Source",
"Included",
"-",
"-",
"-",
"Included",
"-",
"-",
"-",
"-",
"Included",
],
"RedPajamaV1": [
"5",
"1 Source",
"Included",
"",
" ",
"",
"Included",
"-",
"-",
"-",
"Included",
"Included",
],
"The Pile": [
"0.6% of 74",
"4 Sources",
"English Only",
"Included",
"Included",
"Included",
"Included",
"Included",
"Included",
"Included",
"Included",
"Included",
],
}
)
styled_table = (
new_dataset_comparison1.style.applymap(
lambda _: "background-color: #E1EEDB", # Green background for col 1
subset=pd.IndexSlice[:, "TxT360"],
)
.applymap(
lambda _: "background-color: white", # White background for all other columns
subset=pd.IndexSlice[
:, new_dataset_comparison1.columns.difference(["TxT360"])
], # Apply to all columns except "TxT360"
)
.set_properties(
**{
"text-align": "left", # Left the text in each cell
"padding": "10px", # Add padding for better readability
"word-wrap": "break-word", # Ensure text wraps within cells
}
)
.hide(axis="index") # Hide the row index
)
# Use _repr_html_() method to get the HTML representation of the styled DataFrame
table_html = styled_table._repr_html_()
# table_html = dataset_comparison1.to_html(index=False, border=0)
# new_table_div_1 = Div(NotStr(table_html), style="margin: 40px;")
new_table_div_1 = Div(
NotStr(table_html),
style="display: flex; justify-content: center; align-items: center; width: 100%; max-width: 100%; height: auto; overflow-x: auto;"
)
dataset_comparison1 = pd.DataFrame(
{
"Dataset": [
"TxT360",
"FineWeb",
"RefinedWeb",
"RedPajama-v2",
"C4",
"Dolma",
"RedPajama-v1",
"The Pile",
],
"CommonCrawl": [
"99 Snapshots",
"96 Snapshots",
"90 Snapshots",
"84 Snapshots",
"1 Snapshots",
"24 Snapshots",
"5 Snapshots",
"0.6% of 74 Snapshots",
],
"Papers": [
"5 Sources",
"-",
"-",
"-",
"-",
"1 Source",
"1 Source",
"4 Sources",
],
"Wikipedia": [
"310+ Languages",
"-",
"-",
"-",
"-",
"what does a check mark mean?",
"what does a check mark mean?",
"English Only",
],
"FreeLaw": [
"Included",
"-",
"-",
"-",
"-",
"-",
"-",
"Included",
],
"DM Math": [
"Included",
"-",
"-",
"-",
"-",
"-",
"-",
"Included",
],
"USPTO": [
"Included",
"-",
"-",
"-",
"-",
"-",
"-",
"Included",
],
}
)
# Apply table styling: Light green for the header, alternating white and light grey for rows
styled_table = (
dataset_comparison1.style.set_properties(
**{"background-color": "#E1EEDB"},
subset=pd.IndexSlice[0, :], # Row 0 with a light green background
)
.apply(
lambda x: [
"background-color: #E1EEDB" # Green background for row 0
if i == 0
else "background-color: rgb(237, 242, 251)" # Blue background for other rows
for i in range(len(x))
],
axis=0,
)
.hide(axis="index")
) # Hide the row index
# Use _repr_html_() method to get the HTML representation of the styled DataFrame
table_html = styled_table._repr_html_()
# table_html = dataset_comparison1.to_html(index=False, border=0)
table_div_1 = Div(NotStr(table_html), style="margin: 40px;")
dataset_comparison2 = pd.DataFrame(
{
"Dataset": [
"TxT360",
"FineWeb",
"RefinedWeb",
"RedPajama-v2",
"C4",
"Dolma",
"RedPajama-v1",
"The Pile",
],
"PG-19": [
"Included",
"-",
"-",
"-",
"-",
"Included",
"Included",
"Included",
],
"HackerNews": [
"Included",
"-",
"-",
"-",
"-",
"-",
"-",
"Included",
],
"Ubuntu IRC": [
"Included",
"-",
"-",
"-",
"-",
"-",
"-",
"Included",
],
"EuroParl": [
"Included",
"-",
"-",
"-",
"-",
"-",
"-",
"Included",
],
"StackExchange": [
"Included",
"-",
"-",
"-",
"-",
"-",
"Included",
"Included",
],
"Code": [
"- what is this?",
"-",
"-",
"-",
"-",
"Included",
"Included",
"Included",
],
}
)
# Apply table styling: Light green for the header, alternating white and light grey for rows
styled_table = (
dataset_comparison2.style.set_properties(
**{"background-color": "#E1EEDB"},
subset=pd.IndexSlice[0, :], # Row 0 with a light green background
)
.apply(
lambda x: [
"background-color: #E1EEDB"
if i == 0
else (
"background-color: rgb(237, 242, 251)"
if i % 2 == 0
else "background-color: white"
)
for i in range(len(x))
],
axis=0,
)
.set_table_styles(
[
{"selector": "table", "props": [("margin-left", "auto"), ("width", "100%")]}, # Make table responsive and centered
]
)
.hide(axis="index")
) # Hide the row index
# Use _repr_html_() method to get the HTML representation of the styled DataFrame
table_html2 = styled_table._repr_html_()
# table_html2 = dataset_comparison2.to_html(index=False, border=0)
# table_div_2 = Div(NotStr(table_html2), style="margin: 40px;")
table_div_2 = Div(NotStr(table_html2))
dataset_sources = pd.DataFrame(
{
"Data Source": [
"CommonCrawl",
"Papers",
"Wikipedia",
"Freelaw",
"DM Math",
"USPTO",
"PG-19",
"HackerNews",
"Ubuntu IRC",
"Europarl",
"StackExchange",
],
"Raw Data Size": [
"9.2 TB",
"712 GB",
"199 GB",
"71 GB",
"22 GB",
"45 GB",
"11 GB",
"4.1 GB",
"4.7 GB",
"6.1 GB",
"79 GB",
],
"Token Count": [
"4.83T",
"154.96B",
"35.97B",
"16.7B",
"5.23B",
"4.95B",
"2.63B",
"1.08B",
"1.54B",
"1.96B",
"27B",
],
"Information Cut-Off Date": [
"2024-30",
"Q4 2023",
"-",
"Q1 2024",
"-",
"Q3 2024",
"-",
"Q4 2023",
"Q3 2024",
"-",
"Q4 2023",
],
}
)
# Apply table styling: Light green for the header, alternating white and light grey for rows
styled_table = (
dataset_sources.style.apply(
lambda x: [
"background-color: white"
if i % 2 == 0
else "background-color: rgb(237, 242, 251)"
for i in range(len(x))
],
axis=0,
)
.set_properties(
**{
"text-align": "center", # Center the text in each cell
"padding": "10px", # Add padding for better readability
"word-wrap": "break-word", # Ensure text wraps within cells
}
)
.hide(axis="index") # Hide the row index
)
table_html_data = styled_table._repr_html_()
# Wrap the table in a Div, ensuring it is centered
table_div_data = Div(
NotStr(table_html_data),
# style="margin-left: auto; width: 90%; max-width: 100%; text-align: center; align: center; overflow-x: auto;"
style="display: flex; justify-content: center; align-items: center; width: 100%; max-width: 100%; height: auto; overflow-x: auto;"
)
@app.get("/intro")
def intro():
return Div(
Section(
H2("About TxT360"),
P( B("TL;DR "),
"We introduce ",
A(B("TxT360 (Trillion eXtracted Text),"), href="https://huggingface.co/datasets/LLM360/TxT360"),
" the first dataset to globally deduplicate 99 CommonCrawl snapshots and 14 high-quality data sources from diverse domains (e.g., FreeLaw, PG-19, etc.). The large-scale deduplication process and rich metadata stored enables precise control over data distribution. We demonstrate a simple but effective upsampling recipe that creates a 15+ trillion-token corpus, outperforming FineWeb 15T on several key metrics. With the information, TxT360 empowers pre-trainers to explore more advanced weighting techniques, a feature not commonly available in previous pre-training datasets. Our findings highlight the importance of both high-quality data sources and appropriate weighting for optimal blending in LLM training."
),
P("In line with our 360° open source spirit, we document all detailed steps, reasons of our decisions, detailed statistics, our code (stay tuned!), analysis results and more, in addition to the dataset itself. We hope this can serve as a useful resource for future developers."
),
plotly2fasthtml(all_eval_res_figs["MMLU"]),
P(
"Building on top of the prior studies on pre-training data",
D_cite(bibtex_key="refinedweb"),
D_cite(bibtex_key="fineweb"),
D_cite(bibtex_key="c4"),
D_cite(bibtex_key="muennighoff2023scaling"),
D_cite(bibtex_key="dolma"),
", TxT360 carefully implements data processing steps including extraction, filtering, deduplication, personally identifiable information removal, and other steps. Unlike DCLM",
D_cite(bibtex_key="dclm"),
"and RedPajama V2,",
D_cite(bibtex_key="redpajama-v2"),
"we also hope to provide a dataset at this scale that is ready to go, without requiring further filtering."
),
P(
B("How to Read this Blog Post?"),
P("This document contains all the details and is lengthy. We recommend readers to use the Table of Contents to jump to the appropriate sections. The post might also be slightly too long for mobile reading (sorry!). At each top level section, we provided a quick guide for the content. We also recommend readers to consider this post as a reference for some high level statistics related to pre-training datasets."),
P("Advanced blog navigation elements are available on laptops and larger viewing windows."),
),
id="section11",
),
Section(
H2("Why TxT360"),
P(
"In this year we have seen excellent datasets released by the community. Among those, most datasets focus on one source (e.g., crawled websites, code bases, papers). However, it is not trivial to combine these sources together due to the potential duplication across them. TxT360 is the first dataset to combine most of sources commonly used in pretraining."
),
new_table_div_1,
# table_div_1,
# table_div_2,
P(
"In LLM pretraining, it is common to combine all possible text sources due to the Scaling Law. Crawled web pages are included to provide a vast quantity of data which can cover long tail and diverse information, while curated datasets such as Wikipedia are also used, which often provide the 'deep-dive' domain information. By integrating the reach of web data with the quality of curated sources, TxT360 meets and surpasses the rigorous standards required for state-of-the-art LLM pre-training."
),
P(
"** TxT360 does not include very specific domains such as code and math. This decision was made due to the perceived low duplication code with other sources, and the different logic required to build those datasets. We leave that to future work and recommend users refer to existing projects such as Stack V2",
D_cite(bibtex_key="lozhkov2024starcoder2stackv2"),
".",
),
# P("Table 2: Basic TxT360 Statistics."),
# table_div_data,
id="section12",
),
Section(
H2("Our Approach"),
P(
"To produce TxT360, a comprehensive data processing pipeline was designed to account for the nuances of both web and curated datasets. The pipeline presents a unified framework for processing both data types, making it convenient and easily adaptive for users to revise and fine-tune the pipeline for their own use cases."
),
P(
"Web datasets are inherently noisy and varied. The TxT360 pipeline implements sophisticated filtering and deduplication techniques to clean and remove redundancies while preserving data integrity."
),
P(
"Curated datasets are typically structured and consistently formatted, but also can cause troubles with their own special formatting preferences. TxT360 filters these sources with selective steps to maintain their integrity while providing seamless integration into the larger dataset. Both data source types are globally deduplicated together resulting in ~5T tokens of high-quality data. The table below shows the source distribution of TxT360 tokens. ",
B("Note that we do not recommend to use the raw distribution of the deduplicated dataset, a simple recipe is provided in the studies section."),
),
table_div_data,
P(
"We provide details and context for the choices behind TxT360 in the respective Common Crawl Data Processing and Curated Source Processing section. A deep dive describing the deduplication process can be found in the Shared Processing Steps section."
),
# Img(src="images/pipeline.png", height="300", width="600"),
# P(
# "Figure 1: Data processing pipeline. All the steps are adopted for processing web data while the yellow blocks are adopted for processing curated sources."
# ),
id="section13",
),
id="inner-text",
)
rt("/update/{target}")(data_viewer.update)
serve()
|