File size: 64,315 Bytes
e137e27 27b61df 3396b8b e137e27 41b9932 715785a 41b9932 715785a 41b9932 715785a e137e27 0e26631 41b9932 715785a 41b9932 0e26631 8981902 0e26631 5b83110 0e26631 5b83110 0e26631 e137e27 146aa07 e137e27 41b9932 e04322e 32c6d51 146aa07 e137e27 64c513d e04322e b8195ee 6084136 a810b7b e137e27 6084136 a810b7b e137e27 88c0211 6084136 88c0211 e137e27 88c0211 e137e27 88c0211 e137e27 a810b7b 88c0211 073687e 88c0211 073687e e137e27 88c0211 e137e27 88c0211 e137e27 88c0211 e137e27 88c0211 e137e27 88c0211 e137e27 88c0211 e137e27 88c0211 e137e27 6263148 e137e27 aa13e37 e137e27 aa13e37 6263148 e137e27 aa13e37 e137e27 aa13e37 6263148 e137e27 aa13e37 e137e27 aa13e37 e137e27 aa13e37 e137e27 aa13e37 e137e27 aa13e37 465a4f0 aa13e37 e137e27 26832b9 e137e27 aa13e37 465a4f0 aa13e37 465a4f0 aa13e37 6263148 e137e27 aa13e37 e137e27 7e7a96b e137e27 ae1d7f9 aa13e37 ae1d7f9 aa13e37 ae1d7f9 e137e27 ae1d7f9 aa13e37 ae1d7f9 e137e27 ae1d7f9 e137e27 be782f3 e137e27 be782f3 e137e27 be782f3 e137e27 be782f3 e137e27 ae1d7f9 aa13e37 ae1d7f9 e137e27 ae1d7f9 aa13e37 ae1d7f9 e137e27 ae1d7f9 aa13e37 ae1d7f9 aa13e37 ae1d7f9 aa13e37 ae1d7f9 aa13e37 ae1d7f9 e137e27 ae1d7f9 e137e27 7e7a96b e137e27 26832b9 e137e27 6263148 e137e27 7e7a96b ae1d7f9 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 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 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342 1343 1344 1345 1346 1347 1348 1349 1350 1351 1352 1353 1354 1355 1356 1357 1358 1359 1360 1361 1362 1363 1364 1365 1366 1367 1368 1369 1370 1371 1372 1373 1374 1375 1376 1377 1378 1379 1380 1381 1382 1383 1384 1385 1386 1387 1388 1389 1390 1391 1392 1393 1394 1395 1396 1397 1398 1399 1400 1401 1402 1403 1404 1405 1406 1407 1408 1409 1410 1411 1412 1413 1414 1415 1416 1417 1418 1419 1420 1421 1422 1423 1424 1425 1426 1427 1428 |
from fasthtml.common import *
from fasthtml.components import *
import json
import random
import string
from rich import print
import jsonlines
from data.url_blocklist import urls_high_matches, urls_false_positives
from data.non_web_urls import non_web_urls
from fasthtml.components import D_code
import pandas as pd
data_filtering_table_data = pd.DataFrame(
{
"Dataset": [
"TxT360",
"FineWeb",
"RefinedWeb",
"RedPajamaV2",
"C4",
"Dolma",
"RedPajamaV1",
"The Pile",
],
"Data Reading": [
"warc",
"warc",
"warc",
"wet",
"wet",
"warc",
"wet",
"warc",
],
"Text Extraction": [
"trafilatura",
"trafilatura",
"trafilatura",
"n/a",
"n/a",
"?",
"n/a",
"jusText",
],
"URL Filtering": [
"Yes",
"Yes",
"Yes",
"Yes",
"No",
"No",
"No",
"No",
],
"Language Identification": [
"fastText",
"fastText",
"fastText",
"fastText",
"langdetect",
"fastText",
"fastText",
"pycld2",
],
"Line Removal": [
"Yes",
"Yes",
"Yes",
"Yes",
"Yes",
"Yes",
"No",
"No",
],
"PII Filtering": [
"Yes",
"Yes",
"No",
"No",
"No",
"Yes",
"No",
"No",
],
"Exact Deduplication": [
"Bloom Filter",
"n/a",
"ExactSubStr",
"Bloom Filter",
"n/a",
"Bloom Filter",
"n/a",
"n/a",
],
"Fuzzy Deduplication": [
"Global",
"Local",
"Local",
"Local",
"Local",
"Local",
"Local",
"Global",
],
}
)
table_html_filter_data = data_filtering_table_data.to_html(index=False, border=0)
table_div_filter_data = Div(NotStr(table_html_filter_data), style="margin: 40px;")
qf_filtering_table_data = pd.DataFrame(
{
"Dataset": [
"TxT360",
"FineWeb",
"RefinedWeb",
"RedPajamaV2",
"C4",
"Dolma",
"RedPajamaV1",
"The Pile",
],
"QF: ML-based": [
"No",
"No",
"No",
"Yes",
"No",
"No",
"Yes",
"Yes",
],
"QF: Repition-based": [
"Yes",
"Yes",
"Yes",
"Yes",
"No",
"Yes",
"No",
"No",
],
"QF: Correction-based": [
"Yes",
"Yes",
"Yes",
"No",
"No",
"No",
"No",
"No",
],
"QF: Gopher Rules": [
"Yes",
"Yes",
"Yes",
"Yes",
"No",
"Yes",
"No",
"No",
],
"QF: C4 Rules": [
"Yes",
"Yes",
"Yes",
"Yes",
"Yes",
"Yes",
"No",
"No",
],
}
)
table_html_qf_filter_data = qf_filtering_table_data.to_html(index=False, border=0)
table_div_qf_filter_data = Div(NotStr(table_html_qf_filter_data), style="margin: 40px;")
def DVS(
left,
header,
):
col1 = Div(
Pre(
json.dumps(left, indent=4, ensure_ascii=False),
style="white-space: pre-wrap; word-break: break-all;",
),
style="float: left; overflow-x: auto;",
)
data_display = Div(
col1,
style="overflow: auto; clear: both; height: 200px; border: 1px solid #ccc; padding: 20px;",
)
return Div(H3(header), data_display, style="margin-top: 10px;")
def DV(
left_file,
doc_id,
header,
target: str = None,
):
if target is None:
target = "".join(random.choices(string.ascii_lowercase, k=8))
if left_file.endswith("jsonl"):
left = [x for x in jsonlines.open(left_file)]
else:
left = json.load(open(left_file, encoding="utf-8"))
max_doc_id = len(left) - 1
slider = Input(
type="range",
name=f"doc_id_{target}",
min="0",
max=str(max_doc_id),
value=str(doc_id),
hx_get=f"/webdata/{target}",
hx_target=f"#{target}",
hx_trigger="change",
hx_swap="innerHTML",
hx_vals=json.dumps({"left_file": f"{left_file}", "header": f"{header}"}),
)
form = Div(
H3(header),
Label(
"Data sample: ", slider, f"{doc_id} of {max_doc_id}", cls="plotly_slider"
),
cls="plotly_input_container",
style="padding: 20px;",
)
col1 = Div(
Pre(
json.dumps(left[doc_id], indent=4, ensure_ascii=False),
style="white-space: pre-wrap; word-break: break-all;",
),
style="float: left; overflow-x: auto;",
)
data_display = Div(
col1,
style="overflow: auto; clear: both; height: 600px; border: 1px solid #ccc; padding: 20px;",
)
return Div(form, data_display, style="margin-top: 10px;", id=target)
def DV2(
left_file,
right_file,
doc_id,
target: str = None,
):
if target is None:
target = "".join(random.choices(string.ascii_lowercase, k=8))
left = json.load(open(left_file, encoding="utf-8"))
right = json.load(open(right_file, encoding="utf-8"))
max_doc_id = len(left) - 1
slider = Input(
type="range",
name=f"doc_id_{target}",
min="0",
max=str(max_doc_id),
value=str(doc_id),
hx_get=f"/webdata/{target}",
hx_target=f"#{target}",
hx_trigger="change",
hx_swap="innerHTML",
hx_vals=json.dumps(
{"left_file": f"{left_file}", "right_file": f"{right_file}"}
),
)
form = Div(
Label(
"Data sample: ", slider, f"{doc_id} of {max_doc_id}", cls="plotly_slider"
),
cls="plotly_input_container",
style="padding: 20px;",
)
col1 = Div(
H3("Raw format", style="margin-top: 0px;"),
Pre(
json.dumps(left[doc_id], indent=4, ensure_ascii=False),
style="white-space: pre-wrap; word-break: break-all;",
),
style="width: 48%; float: left; overflow-x: auto;",
)
col2 = Div(
H3("Extracted format", style="margin-top: 0px;"),
Pre(
json.dumps(right[doc_id], indent=4, ensure_ascii=False),
style="white-space: pre-wrap; word-break: break-all;",
),
style="width: 48%; float: right; overflow-x: auto;",
)
data_display = Div(
col1,
col2,
style="overflow: auto; clear: both; height: 600px; border: 1px solid #ccc; padding: 20px;",
)
return Div(form, data_display, style="margin-top: 10px;", id=target)
def update(target: str, request):
params = request.query_params
print(params)
doc_id = int(params.get(f"doc_id_{target}", 3))
left_file = params.get("left_file")
right_file = params.get("right_file")
if left_file and right_file:
return (
DV2(
left_file,
right_file,
doc_id,
target,
),
)
else:
return DV(
left_file,
doc_id,
params.get("header"),
target,
)
dolma311 = """
words = text.split()
word_count = len(words)
character_count = sum(len(word) for word in words)
...
lines = text.split("\\n")
line_count = len(lines)
...
line_counts = Counter(lines)
attrs.fraction_of_duplicate_lines = sum(count for line, count in line_counts.items() if count > 1) / max(
line_count, 1
)
attrs.fraction_of_characters_in_duplicate_lines = sum(
len(line) * count for line, count in line_counts.items() if count > 1
) / max(character_count, 1)
"""
def web_data():
return Div(
Div(
Ul(
Li(
A(
"Raw Documentation",
href="https://drive.google.com/drive/folders/1mIJ-Zx8tRhohFdj4ByMToNz1u_9Saa8W?usp=drive_link",
)
),
Li(
A(
"Github link of Web Data Pipeline",
href="https://github.com/CIAI-LLM/WebDataProcessing.git",
)
),
),
style="""
background-color: #d4edda; /* Light green background */
border: 1px solid #c3e6cb; /* Green border */
border-radius: 5px;
padding: 15px 15px 0px 15px;
""",
),
Div(
P(
"To generate a high-quality dataset from large-scale webpages, we have investigated the processing steps used by the community and made our choices based on careful manual inspection. Starting from ",
A("Common Crawl", href="https://commoncrawl.org/"),
", our process can be summarized as five main steps: document preparation, line-level removal, document-level filtering, deduplication and PII removal.",
),
style="margin-top: 20px;",
),
H2("Web Data Processing Summary"),
P("The following section provides explicit details covering the reasoning and decisions behind each of the filters we applied. The table below provides a high-level comparison of TxT360's filtering compared to other commonly used pretraining datasets."),
table_div_filter_data,
P("ADD EXPLAINER TEXT ABOUT THE QUALITY FILTERS"),
table_div_qf_filter_data,
P("Our filtering rate is illustrated below. Before deduplication, our filtering rate is comparable to RefinedWeb. During global deduplication, we removed approximately 85.89% of the data, significantly higher than previous works, indicating a large number of duplicates across dumps. "),
Img(src="images/filter_rate.jpg", height = "300", width = "600" ),
P("Note: All percentages are based on the number of documents. The gray bars represent the relative percentages of removed documents at each step, while the colorful bars represent the percentages of retained documents relative to the total number of documents in the raw Common Crawl."),
P("We also adopt rules from RefinedWeb [1] to remove lines if they satisfy any of the following criteria:"),
Ul(
Li("the line is only composed of uppercase characters", style = "margin-bottom: 5px"),
Li("the line is only composed of numerical characters", style = "margin-bottom: 5px"),
Li("the line matches the pattern “r'^\d+\s+likes$", style = "margin-bottom: 5px"),
Li("the line only contains one word.", style = "margin-bottom: 5px"),
),
P("We summarize other statistics-based rules originated from Gopher [7] in this section. The statistics can be used include:"),
Ul(
Li("the word count in the document", style = "margin-bottom: 5px"),
Li("the mean word length", style = "margin-bottom: 5px"),
Li("the number of sentences", style = "margin-bottom: 5px"),
Li("the symbol-to-word ratio", style = "margin-bottom: 5px"),
Li("the fraction of alphabetic words", style = "margin-bottom: 5px"),
Li("and the number of stop words", style = "margin-bottom: 5px"),
),
P("Specifically, we remove any document which satisfies any of the following criteria:"),
Ul(
Li("it contains less than 50 words or more than 100,000 words", style = "margin-bottom: 5px"),
Li("its mean word length is outside the range of 3 to 10", style = "margin-bottom: 5px"),
Li("it contains less than 3 sentences", style = "margin-bottom: 5px"),
Li("its symbol-to-word ratio is greater than 0.1", style = "margin-bottom: 5px"),
Li("the words that contain at least one alphabetic character are less than 80% of the whole words", style = "margin-bottom: 5px"),
Li("it contains less than two of the stop words (the, be, to, of, and, that, have, with", style = "margin-bottom: 5px"),
),
P("Following C4, we remove any page where the phrase “lorem ipsum” appears since some pages have placeholder “lorem ipsum” text."),
H3("1. Document Preparation"),
H4("1.1 Text Extraction"),
P("""
Common Crawl provides webpage texts via two formats: WARC (Web ARChive format) and WET (WARC Encapsulated Text).
WARC files contain the raw data from the crawl, which store the full HTTP response and request metadata.
WET files contain plaintexts extracted by Common Crawl. In line with previous works ([1], [2], [3], [4]),
we found WET files to include boilerplate content like navigation menus, ads, and other irrelevant texts.
Accordingly, our pipeline starts from raw WARC files, reads with the warcio library, and extracts texts using trafilatura.
"""),
P("We directly read WARC files instead of WET files and extracted text using Trafilatura. Similar to RefinedWeb, we avoid using Machine Learning (ML)-based metrics for filtering documents to prevent bias introduced by ML models. Importantly, we apply global deduplication across the entire dataset, whereas previous works only use local deduplication. Note that although The Pile also employed global deduplication on its web data (Pile-CC), this accounted for just 0.6\% of 74 snapshots."),
Details(
Summary("Text Extraction Examples"),
DV2("data/sample_wet.json", "data/sample_warc.json", 3),
),
#DV2("data/sample_wet.json", "data/sample_warc.json", 3),
H4("1.2 Language Identification"),
P("""
After text extraction, the non-English texts are then filtered out by fastText language identifier with a threshold of 0.65.
This step removes over 60% of the whole data.
"""),
Details(
Summary("Non-English Documents"),
DV("data/sample_non_en.json", 3, "Sample documents that are classified as non-English"),
),
#DV("data/sample_non_en.json", 3, "Sample documents that are classified as non-English"),
Details(
Summary("English Documents Scoring Lower than 0.65"),
DV("data/sample_en_low.json", 3, "Sample documents that are classified as English but with score less than 0.65"),
),
H4("1.3 URL Filtering"),
P("""
Following RefinedWeb [3], we use a manually inspected URL blocklist to filter fraudulent and/or adult websites.
We also exclude our high-quality curated data from it to avoid duplication.
"""),
H5("1.3.1 URL Blocklist"),
P("""
Following RefinedWeb [3], we applied manual inspection on the UT1 blocklist to reduce false positives like news
articles, sex education, technical blogs, etc. Specifically, we randomly took 903M URLs and matched them with
4.6M domain names in the UT1 blocklist. 24 URL domains were detected with more than 4k matches, which are shown below.
"""),
Details(
Summary("24 URL domains with more than 4k matches"),
DVS(urls_high_matches, "24 URL domains with more than 4k matches"),
),
P("""
We manually removed the following 6 domains from the UT1 blocklist so that they will not be removed from our dataset.
"""),
Details(
Summary("6 url domains that are removed from the blocklist"),
DVS(urls_false_positives, "6 url domains that are removed from the blocklist"),
),
Details(
Summary("Sample documents whose urls are blocked by the refined url blocklist"),
DV(
"data/bad_url_doc.jsonl",
3,
"Sample documents whose urls are blocked by the refined url blocklist",
),
),
H5("1.3.2 Excluded High Quality Sources"),
P("""
To avoid duplication with our high-quality curated datasets, we exclude the following domains from our dataset.
"""),
Details(
Summary("curated url domains that are excluded from our dataset"),
DVS(
non_web_urls,
"curated url domains that are excluded from our dataset",
),
),
Details(
Summary("Sample documents whose urls are in our curated url domain list"),
DV("data/sample_url_exclusion.json", 0, "Sample documents whose urls are in our curated url domain list"),
),
H3("2. Line-Level Removal"),
P("""
Before computing the quality signals that can be used for filtering low-quality documents, we perform the line-level
removal to remove low-quality lines so that the final quality signals align with our final kept texts.
"""),
H4("Terminal Punctuation"),
P("""
The terminal punctuation has been used in C4 [5] and Dolma [6] to remove lines that do not end with a terminal
punctuation mark (i.e., “.”, “?”, “!”, or “"”). However, we found it could be too aggressive to remove these
lines, especially when using a better text extraction tool “trafilatura”. For instance, in the file
CC-MAIN-20230126210844-20230127000844-00000.warc.jsonl, the terminal punctuation rule led to the removal
of 56,292 additional lines, resulting in the complete exclusion of 2,203 documents from a total of 13,560
documents (16.25%). Accordingly, we choose to not use terminal punctuation as a signal to remove lines.
"""),
Details(
Summary("Sample documents with lines that are removed by the rule of terminal punctuation"),
DV(
"data/sample_terminal_punc.json",
0,
"Sample documents with lines that are removed by the rule of terminal punctuation",
),
),
H4('2.1 Word "Javascript"'),
P("""
In C4 [5], the authors remove any line with the word "Javascript" since they found that many of the scraped
pages contained warnings stating that Javascript should be enabled. However, this filtering strategy is too
strict, which will filter out many lines that are really talking about “Javascript”. In our pipeline, we
propose to refine the strategy by adding one more keyword to the word "javascript" to avoid false positives.
The additional keyword could be any one of “enable” / “disable” / “require” / “activate” / “browser”.
"""),
Details(
Summary("Sample documents that are removed by original C4 javascript rule but are kept after our refinement"),
DV(
"data/sample_java.jsonl",
0,
"Sample documents that are removed by original C4 javascript rule but are kept after our refinement",
),
),
H4("2.2 Other Rules from RefinedWeb"),
P("""
We also adopt rules from RefinedWeb [3] to remove lines if they satisfy any of the following criteria:
- The line is only composed of uppercase characters,
- The line is only composed of numerical characters,
- The line matches the pattern “r'^\\d+\\s+likes$'”,
- The line contains only one word.
"""),
Details(
Summary("Sample documents with lines that are removed by the RefinedWeb rules"),
DV(
"data/sample_refinedweb_line.json",
0,
"Sample documents with lines that are removed by the RefinedWeb rules",
),
),
H4("2.3 Toxic Lines"),
P("""
When doing manual inspection on the data, we found that there are some adult ads in the beginning or end of the
document (with a sample shown below), which are hard to remove via document-level filtering strategies. Inspired
by this, we develop line-level detoxification using a bad word list from LDNOOBW (+ rule: word length < 10 + the
line is in the first 3 lines or in the last 3 lines) to remove toxic lines. Specifically, we do not only consider
the bad words from English but also consider the bad words from other languages.
"""),
Details(
Summary("Sample documents with toxic lines"),
DVS(
json.load(open("data/toxic_lines.json")),
"Sample documents with toxic lines",
),
),
H3("3. Document-Level Filtering"),
P("""
In this section, we introduce all the quality signals that we have used to filter out low-quality documents.
Overview of all the quality signals that are used for filtering."""),
Details(
Summary("Overview of all the quality signals that are used for filtering"),
DVS(
json.load(open("data/all_signals.json")),
"Overview of all the quality signals that are used for filtering",
),
),
P("""Similar to previous sections, we will present sample documents filtered out by the given quality signals.
Most of these quality signals were initially introduced by Gopher [2] and subsequently adopted by later
studies ([3], [6], [4]). However, we observed that, despite following the same descriptions, the implementation
of each quality signal can vary significantly among different dataset pipelines, resulting in disparate
outcomes for the same quality signals.
In our pipeline, we referenced earlier implementations that were publicly available such as Dolma [6], DataTrove [4],
and RedPajama V2 [7], selecting the most suitable method based on manual inspections.
"""),
H4("3.1 Repetition-based Heuristics"),
P("""
Due to crawling errors or low-quality sources, many documents contain repeated sequences. In line with previous
work ([2], [3], [6]), we choose to remove any document with excessive line, paragraph, or n-gram repetitions.
"""),
H5("3.1.1 Fraction of (Characters in) Repeated Lines"),
P("""
Following Gopher [2], we remove documents containing many short duplicate passages, as well as those with few,
but longer duplicate passages. To achieve this goal, we calculate over the document both the fraction of passages
that are duplicates, and the fraction of characters contained within those duplicated passages.
"""),
Details(
Summary("Implementations from Dolma"),
D_code("""
words = text.split()
word_count = len(words)
character_count = sum(len(word) for word in words)
...
lines = text.split("\n")
line_count = len(lines)
...
line_counts = Counter(lines)
attrs.fraction_of_duplicate_lines = sum(count for line, count in line_counts.items() if count > 1) / max(
line_count, 1
)
attrs.fraction_of_characters_in_duplicate_lines = sum(
len(line) * count for line, count in line_counts.items() if count > 1
) / max(character_count, 1)
""", block="block", language="python"),
),
Details(
Summary("Implementations from DataTrove"),
D_code("""
def find_duplicates(x: list[str]) -> tuple[int, int]:
unique_x = set()
duplicate_chars = 0
duplicate_elements = 0
for element in x:
if element in unique_x:
duplicate_chars += len(element)
duplicate_elements += 1
else:
unique_x.add(element)
return duplicate_elements, duplicate_chars
...
self.paragraph_exp = re.compile(r"\n{2,}")
self._line_splitter = re.compile("\n+")
...
paragraphs = self.paragraph_exp.split(text.strip())
paragraphs_duplicates, char_duplicates = find_duplicates(paragraphs)
if self.dup_para_frac and paragraphs_duplicates / len(paragraphs) > self.dup_para_frac:
return False, "dup_para_frac"
if self.dup_para_char_frac and char_duplicates / len(text) > self.dup_para_char_frac:
return False, "dup_para_char_frac"
lines = self._line_splitter.split(text)
line_duplicates, char_duplicates = find_duplicates(lines)
if self.dup_line_frac and line_duplicates / len(lines) > self.dup_line_frac:
return False, "dup_line_frac"
if self.dup_line_char_frac and char_duplicates / len(text) > self.dup_line_char_frac:
return False, "dup_line_char_frac"
""", block="block", language="python"),
),
P("""
After evaluating the implementations of Dolma and DataTrove (note: RedPajama V2 does not implement these two quality
signals), we have made the following decisions:
"""),
H5("Passage Separation"),
P("""
Our manual review of the data revealed that documents extracted using trafilatura do not feature more than one newline
symbol separating passages. Testing the splitting pattern "\\n(2,)" on 10,000 sample documents resulted in no more than
one split. Consequently, we decided to disregard the distinction between lines and paragraphs in our implementation,
opting instead to use a single newline symbol to segment the text into passages.
"""),
H5("First Occurrence"),
P("""
In line with DataTrove's implementation, we chose to exclude the first occurrence. This more conservative strategy
helps retain a larger number of documents.
"""),
H5("Character Count"),
P("""
We adjusted the method in Dolma for counting characters within lines by excluding whitespace. This modification
ensures consistency with the overall document character count calculation.
"""),
H5("Our Implementation"),
Details(
Summary("TxT360 Implementation"),
D_code("""
words = text.split()
word_count = len(words)
character_count = sum(len(word) for word in words)
...
lines = text.split("\n")
line_count = len(lines)
line_counts = Counter(lines)
attrs.fraction_of_duplicate_lines = (
sum((count - 1) for line, count in line_counts.items() if count > 1) / line_count
)
attrs.fraction_of_characters_in_duplicate_lines = (
sum(sum(len(w) for w in line.split()) * (count - 1) for line, count in
line_counts.items() if count > 1) / character_count
""", block="block", language="python"),
),
Details(
Summary("Sample documents filtered by excessive line repetitions / characters in repeated lines"),
DV(
"data/repeat_line_frac.jsonl",
0,
"Sample documents filtered by excessive line repetitions / characters in repeated lines",
),
),
H5("3.1.2 Fraction of Characters in the Most Common N-grams (n=2,3,4)"),
P("""
Following Gopher [2], we remove documents with a high portion of n-grams. For each n ∈ (2, 3, 4), we calculate the
fraction of characters contained within the most frequently-occurring n-gram.
"""),
Details(
Summary("Implementations from Dolma"),
D_code("""
def all_ngram_counts(words) -> List[Tuple[int, CounterType[Tuple[str, ...]]]]:
return [(n, Counter(list(zip(*[words[i:] for i in range(n)])))) for n in range(2, 11)]
...
all_counts = all_ngram_counts(words)
count_most_common_ngrams = (2, 3, 4)
for n, ngram_counts in all_counts:
if not ngram_counts:
continue
if n in count_most_common_ngrams:
most_common_ngram, count = ngram_counts.most_common(1)[0]
value = count * sum(len(w) for w in most_common_ngram) / max(character_count, 1)
attrs.fraction_of_characters_in_most_common_ngram.append((n, value))
""", block="block", language="python"),
),
Details(
Summary("Implementations from RedPajama-V2"),
D_code("""
class Base_RPS_Frac_Chars_In_Top_NGram(RPSBase): # noqa
## Base class for calculating the fraction of characters in the top N-gram. This operates on the lower-cased, punctation removed content.
NGRAM_SIZE: int = None
__slots__ = []
def __call__(self, document: Document) -> SignalType:
if self.NGRAM_SIZE is None:
raise NotImplementedError(
"NGRAM_SIZE must be set in the subclass"
)
# get the most common ngram
most_common_ngram = Counter(
# fetch the ngrams from the document if they exist, otherwise
# compute them
getattr(document, f"norm_self.NGRAM_SIZEgrams", None)
or
form_ngrams(iter(document.normalized_words), self.NGRAM_SIZE)
).most_common(1)
if len(most_common_ngram) == 0:
return [(0, len(document), 0.0)]
ngram, count = most_common_ngram[0]
if count <= 1:
return [(0, len(document), 0.0)]
total_chars = sum(len(w) for w in document.normalized_words)
score = sum(len(w) for w in ngram) * count / total_chars
score = round(score, PRECISION)
return [(0, len(document), score)]
""", block="block", language="python"),
),
Details(
Summary("Implementations from DataTrove"),
D_code("""
def get_n_grams(words: list[str], n: int) -> list[str]:
return [" ".join(words[i : i + n]) for i in range(len(words) - n + 1)]
def find_top_duplicate(x: list[str]) -> int:
counter = Counter()
for element in x:
counter[element] += 1
top_n_gram = counter.most_common(1)[0]
return len(top_n_gram[0]) * top_n_gram[1]
...
for n, n_frac in self.top_n_grams:
n_grams = get_n_grams(words, n)
if not n_grams:
continue
top_char_length = find_top_duplicate(n_grams)
if top_char_length / len(text) > n_frac:
return False, f"top_n_gram"
""", block="block", language="python"),
),
P("""
There are almost no contradictions between above implementations of fractions of characters in the most common
n-gram. The main process involves counting the occurrences of each n-gram and selecting the most common one. The
fraction is then determined by dividing the number of characters in the most common n-gram by the total number of
characters. One minor difference is that Dolma and DataTrove calculate the fraction of the most common n-gram even
if it only appears once, while RedPajama V2 skips this case.
We choose to follow Dolma and DataTrove by not skipping cases where the most common n-gram occurs only once.
In practice, documents affected by this rule — where the most common n-gram exceeds a given threshold and occurs
only once — tend to be short.
"""),
Details(
Summary("TxT360 Implementation"),
D_code("""
def all_ngram_counts_new(words) -> List[Tuple[int, CounterType[Tuple[str, ...]]]]:
return [(n, list(zip(*[words[i:] for i in range(n)]))) for n in range(2, 11)]
...
all_counts = all_ngram_counts_new(words)
count_most_common_ngrams = (2, 3, 4)
for n, ngram_counts in all_counts:
if not ngram_counts:
continue
if n in count_most_common_ngrams:
most_common_ngram, count = Counter(ngram_counts).most_common(1)[0]
value = count * sum(len(w) for w in most_common_ngram) / character_count
attrs.fraction_of_characters_in_most_common_ngram.append((n, value))
""", block="block", language="python"),
),
Details(
Summary("Sample documents filtered by the fraction of characters in the most common n-grams (n=2,3,4)"),
DV(
"data/sample_top_ngram.json",
0,
"Sample documents filtered by the fraction of characters in the most common n-grams (n=2,3,4)",
),
),
H5("3.1.3 Fraction of Characters in Duplicated N-grams (n=5,...,10)"),
P("""
Following Gopher [2], we remove documents with a high portion of n-grams. For each n ∈ (5, ..., 10), we calculate the
fraction of characters contained within all duplicate n-grams, taking care not to count characters that occur in
overlapping n-grams more than once.
"""),
Details(
Summary("Implementations from Dolma"),
D_code("""
def all_ngram_counts(words) -> List[Tuple[int, CounterType[Tuple[str, ...]]]]:
return [(n, Counter(list(zip(*[words[i:] for i in range(n)])))) for n in range(2, 11)]
...
all_counts = all_ngram_counts(words)
for n, ngram_counts in all_counts:
if not ngram_counts:
continue
if n in count_most_common_ngrams:
...
else:
ng_char_count = sum(count * sum(len(w) for w in ng) for ng, count in ngram_counts.items())
value = sum(
count * sum(len(w) for w in ng) for ng, count in ngram_counts.items() if count > 1
) / max(ng_char_count, 1)
attrs.fraction_of_characters_in_duplicate_ngrams.append((n, value))
""", block="block", language="python"),
),
Details(
Summary("Implementations from RedPajama-V2"),
D_code("""
class Base_RPS_Frac_Chars_In_Dupe_NGrams(RPSBase): # noqa
## Base class for calculating the fraction of characters in duplicate word N-grams. This operates on the lower-cased, punctation removed content. The function also ensures that characters in overlapping ngrams are only counted once.
NGRAM_SIZE: int = None
__slots__ = []
def __call__(self, document: Document) -> SignalType:
if self.NGRAM_SIZE is None:
raise NotImplementedError(
"NGRAM_SIZE must be set in the subclass"
)
if len(document.normalized_words) < self.NGRAM_SIZE:
return [(0, len(document), 0.0)]
# fetch the ngrams from the document if they exist, otherwise
# compute them
doc_n_grams = (
getattr(document, f"norm_self.NGRAM_SIZEgrams", None)
or
tuple(form_ngrams(
iter(document.normalized_words), self.NGRAM_SIZE
))
)
# keep only ngrams which occur at least twice
ngram_dupes =
ngram for ngram, count in Counter(doc_n_grams).items() if count > 1
duplicated_grams = np.zeros(len(document.normalized_words), dtype=int)
i = 0
for ngram in doc_n_grams:
if ngram in ngram_dupes:
duplicated_grams[i: i + self.NGRAM_SIZE] = 1
i += 1
word_lengths = np.array(list(map(len, document.normalized_words)))
chars_duped = np.sum(word_lengths * duplicated_grams)
total_chars = np.sum(word_lengths)
if total_chars == 0:
return [(0, len(document), 0.0)]
score = float(chars_duped / total_chars)
score = round(score, PRECISION)
return [(0, len(document), score)]
""", block="block", language="python"),
),
Details(
Summary("Implementations from DataTrove"),
D_code("""
def find_all_duplicate(words: list[str], n: int) -> int:
n_words = len(words)
unique = set()
repeated_chars, idx = 0, 0
while idx < n_words - n + 1:
n_gram = "".join(words[idx : idx + n])
if n_gram in unique:
repeated_chars += len(n_gram)
idx += n
else:
unique.add(n_gram)
idx += 1
assert repeated_chars <= len("".join(words))
return repeated_chars
...
for n, n_frac in self.dup_n_grams:
n_duplicates_char = find_all_duplicate(words, n)
if n_duplicates_char / len(text) > n_frac:
return False, f"duplicated_n_grams"
""", block="block", language="python"),
),
P("""
For the computation of fraction of characters in duplicate n-gram, Dolma uses the number of characters in all
n-grams (with overlapping) as the denominator, and uses the number of characters in all duplicated n-grams
(with overlapping) as the numerator."""),
P("""RedPajama V2 uses the number of all characters in (the words of) the document
(without overlapping) as the denominator, and uses the number of characters that are recognized as part of the
duplicate n-gram as the numerator."""),
P("""Datatrove uses the number of all characters in the document (including white
spaces, without overlapping) as the denominator, and uses the number of characters that are recognized as
duplicate n-gram as the numerator. However, there is a mismatch in DataTrove’s calculation, as the number of
characters in the duplicated n-grams excludes white spaces, while the total character count of the document
does not."""),
P("""We decided to use the RedPajama V2 implementation but skip the 1st occurrence of the duplicate n-gram.
"""),
Details(
Summary("TxT360 Implementation"),
D_code("""
def get_dup_ngram_frac(n, doc_n_grams, text):
# fetch the ngrams from the document if they exist, otherwise compute them
# doc_n_grams = list(zip(*[words[i:] for i in range(n)]))
duplicated_grams = np.zeros(len(text.split()), dtype=int)
unique_ngrams = set()
for i, ngram in enumerate(doc_n_grams):
if ngram in unique_ngrams:
duplicated_grams[i: i + n] = 1
else:
unique_ngrams.add(ngram)
word_lengths = np.array(list(map(len, text.split())))
chars_duped = np.sum(word_lengths * duplicated_grams)
total_chars = np.sum(word_lengths)
return float(chars_duped / total_chars)
def all_ngram_counts_new(words) -> List[Tuple[int, CounterType[Tuple[str, ...]]]]:
return [(n, list(zip(*[words[i:] for i in range(n)]))) for n in range(2, 11)]
...
all_counts = all_ngram_counts_new(words)
count_most_common_ngrams = (2, 3, 4)
for n, ngram_counts in all_counts:
if not ngram_counts:
continue
if n in count_most_common_ngrams:
...
else:
score = get_dup_ngram_frac(n, ngram_counts, text)
attrs.fraction_of_characters_in_duplicate_ngrams.append((n, score))
""", block="block", language="python"),
),
Details(
Summary("An example to show the difference between above implementations"),
P("""
Considering n = 5 and the sample sentence:
"word_a word_b word_c word_d word_e word_f word_g word_a word_b word_c word_d word_e word_f word_g word_a word_b word_c"
In Dolma's implementation, there are 13 5-grams in total with 6 duplicated 5-grams. The resulting fraction of characters in duplicate 5-gram is 6/13.
In RedPajama's V2 implementation, there are 17*6 characters in total and 14*6 characters that are contained in duplicate 5-grams. The fraction is 14/17.
In DataTrove's implementation, there are 17*6 + 16(white spaces) characters in total and 10 duplicated 5-grams after excluding the first occurrence. The resulting fraction number is 10*6/(17*6+16).
In our implementation, there are 17*6 characters in total with 10*6 characters that are duplicated after excluding the first occurence. This results in a fraction of 10/17.
"""),
),
H5(
"Sample Documents Filtered by the Fraction of Characters in Duplicated N-grams (n=5,...,10)"
),
Details(
Summary("Sample documents filtered by the fraction of characters in duplicated n-grams (n=5,...,10)"),
DV(
"data/sample_dup_ngram.json",
0,
"Sample documents filtered by the fraction of characters in duplicated n-grams (n=5,...,10)",
),
),
H4("3.2 Line-wise Heuristics"),
P("""
Some line-wise information could also be helpful to distinguish low-quality and high-quality documents. Following
RefinedWeb [3], we remove the document if the corrected lines represent more than 5% of words. In line with previous
works ([2], [3], [6]), we remove the documents if more than 30% of the lines end with an ellipsis or more than
90% of lines start with a bullet point.
"""),
Details(
Summary("Ellipsis Symbol Identification Implemetations"),
P("Dolma: "),
D_code("""
ELLIPSIS_SYMBOLS = ("…")
""", block="block", language="python"),
P("RedPajamaV2: "),
D_code("""
ELLIPSIS_SYMBOLS = ("...", "…")
""", block="block", language="python"),
P("DataTrove: "),
D_code("""
ELLIPSIS_SYMBOLS = ("...", "…")
""", block="block", language="python"),
P("TxT360: "),
D_code("""
ELLIPSIS_SYMBOLS = ("...", "…", "[...]", "[…]")
""", block="block", language="python"),
),
Details(
Summary("Bullet Point Identification Implemetations"),
P("Dolma: "),
D_code("""
BULLET_POINTS = ("*", "-"
""", block="block", language="python"),
P("RedPajamaV2: "),
D_code("""
BULLET_POINT_SYMBOLS = (
"•", # bullet point
"‣", # triangular bullet point
"▶", # black right pointing triangle
"◀", # black left pointing triangle
"◦", # white bullet point
"■", # black square
"□", # white square
"▪", # black small square
"▫", # white small square
"–", # en dash
)
""", block="block", language="python"),
P("DataTrove: "),
D_code("""
BULLET_POINT_SYMBOLS = ("•" , "-")
""", block="block", language="python"),
P("TxT360: "),
D_code("""
BULLET_POINT_SYMBOLS = (
"•", # • bullet point
"‣", # ‣ triangular bullet point
"▶", # ▶ black right pointing triangle
"◀", # ◀ black left pointing triangle
"◦", # ◦ white bullet point
"■", # ■ black square
"□", # □ white square
"▪", # ▪ black small square
"▫", # ▫ white small square
"-", # - en dash
"–", # – dash
"—", # — zh dash
"*", # * star
)
""", block="block", language="python"),
),
Details(
Summary("Sample documents that are filtered out by line-wise heuristics"),
DV(
"data/line_info.json",
0,
"Sample documents that are filtered out by line-wise heuristics",
),
),
H4("3.3 Statistics-based Heuristics"),
P("We summarize other statistics-based rules originated from Gopher [7] in this section. The statistics can be used include:"),
Ul(
Li("the word count in the document", style = "margin-bottom: 5px"),
Li("the mean word length", style = "margin-bottom: 5px"),
Li("the number of sentences", style = "margin-bottom: 5px"),
Li("the symbol-to-word ratio", style = "margin-bottom: 5px"),
Li("the fraction of alphabetic words", style = "margin-bottom: 5px"),
Li("and the number of stop words", style = "margin-bottom: 5px"),
),
P("Specifically, we remove any document which satisfies any of the following criteria:"),
Ul(
Li("it contains less than 50 words or more than 100,000 words", style = "margin-bottom: 5px"),
Li("its mean word length is outside the range of 3 to 10", style = "margin-bottom: 5px"),
Li("it contains less than 3 sentences", style = "margin-bottom: 5px"),
Li("its symbol-to-word ratio is greater than 0.1", style = "margin-bottom: 5px"),
Li("the words that contain at least one alphabetic character are less than 80% of the whole words", style = "margin-bottom: 5px"),
Li("it contains less than two of the stop words (the, be, to, of, and, that, have, with", style = "margin-bottom: 5px"),
),
H5("Word Count"),
Details(
Summary("Implementations from Dolma"),
D_code("""
words = text.split()
word_count = len(words)
""", block="block", language="python"),
),
Details(
Summary("Implementations from RedPajama-V2"),
D_code("""
# the normalized content: lowercased and punctuation removed
self._normalized_content = normalize(content)
self._normalized_words = tuple(self._normalized_content.split())
self._num_normalized_words = len(self._normalized_words)
...
def normalize(
text: str,
remove_punct: bool = True,
lowercase: bool = True,
nfd_unicode: bool = True,
white_space: bool = True
) -> str:
#Normalize the text by lowercasing and removing punctuation.
# remove punctuation
if remove_punct:
text = text.translate(TRANSLATION_TABLE_PUNCTUATION)
# lowercase
if lowercase:
text = text.lower()
if white_space:
text = text.strip()
text = re.sub(r"\s+", " ", text)
# NFD unicode normalization
if nfd_unicode:
text = unicodedata.normalize("NFD", text)
return text
""", block="block", language="python"),
),
Details(
Summary("Implementations from DataTrove"),
D_code("""
words = self.tokenizer.word_tokenize(text)
n_words = len(words)
non_symbol_words = [w for w in words if any(ch not in PUNCTUATION_SET for ch in w)]
n_non_symbol_words_words = len(non_symbol_words)
""", block="block", language="python"),
),
P("""
Both Dolma and RedPajama V2 split texts into words using white spaces and newline symbols. However,
DataTrove employs a tokenizer to split texts into words and ignore punctuations, resulting in a higher
word count compared to simple `text.split()`.
We decided to use simple `len(text.split())` to compute the word count.
"""),
H5("Mean Word Length"),
P("""
There is minimal variation among existing pipeline implementations. We simply compute the mean word length as follows:
"""),
D_code("""
words = text.split()
word_count = len(words)
character_count = sum(len(word) for word in words)
mean_word_length = character_count / word_count
""", block="block", language="python"),
P("""
It's worth noting that Dolma used the median word length instead of the mean in their codes.
"""),
D_code("""
from statistics import median
median_word_length = median(len(word) for word in words)
""", block="block", language="python"),
H5("Number of Sentences"),
P("""
The only publicly available implementation of this quality signal is from RedPajama V2, which uses regular expressions
to split text into sentences.
"""),
Details(
Summary("Implementations from RedPajama-V2"),
D_code("""
class RPS_Doc_Num_Sentences(RPSBase): # noqa
##The number of sentences in the content. This is calculated using the regex r'[^.!?]+[.!?]*'
SENT_PATTERN = re.compile(r'[^.!?]+[.!?]*', flags=re.UNICODE)
__slots__ = ()
def __call__(self, document: Document) -> SignalType:
##count the number of sentences in the content using regex
score = float(len(self.SENT_PATTERN.findall(document.raw_content)))
return [(0, len(document), score)]
""", block="block", language="python"),
),
P("""
However, we found that this approach can mistakenly interpret periods in URLs as sentence endings. To address this,
we opted to use `nltk.tokenize.sent_tokenize` for more accurate sentence splitting.
"""),
Details(
Summary("TxT360 Implementation"),
D_code("""
from nltk.tokenize import sent_tokenize
...
def count_sentences(text):
sentences = sent_tokenize(text)
return len(sentences)
...
attrs.num_of_sentences = count_sentences(text)
""", block="block", language="python"),
),
H5("Symbol to Word Ratio"),
P("""
Following RedPajama-V2 and DataTrove, we use the symbols of ("#", "...", "…").
We calculate the ratio as the number of symbols divided by the total number of words.
"""),
Details(
Summary("Implementations from Dolma"),
D_code("""
SYMBOLS = ("#", "…")
...
attrs.symbol_to_word_ratio = sum(1 for word in words if any(s in word for s in SYMBOLS)) / max(
word_count, 1
)
""", block="block", language="python"),
),
Details(
Summary("Implementations from RedPajama-V2"),
D_code("""
class RPS_Doc_Symbol_To_Word_Ratio(RPSBase): # noqa
##The ratio of symbols to words in the content. This is analogous to
##the signal used in Gopher. Symbols are defined "#", "...", and "…".
SYMBOLS = ("#", "...", "…")
__slots__ = ()
def __call__(self, document: Document) -> SignalType:
num_words = document.num_raw_words
if num_words == 0:
return [(0, len(document), None)]
# count the number of symbols in the content
num_symbols = float(sum(
document.raw_content.count(x) for x in self.SYMBOLS
))
score = num_symbols / num_words
score = round(score, PRECISION)
return [(0, len(document), score)]
""", block="block", language="python"),
),
Details(
Summary("Implementations from DataTrove"),
D_code("""
if self.max_symbol_word_ratio and text.count("#") / n_words > self.max_symbol_word_ratio:
return False, "gopher_too_many_hashes"
if self.max_symbol_word_ratio and (text.count("...") + text.count("…")) / n_words > self.max_symbol_word_ratio:
return False, "gopher_too_many_ellipsis"
""", block="block", language="python"),
),
Details(
Summary("TxT360 Implementation"),
D_code("""
SYMBOLS = ("#", "...", "…")
...
symbol_pattern = re.compile("|".join(re.escape(symbol) for symbol in SYMBOLS))
...
attrs.symbol_to_word_ratio = sum(1 for word in words if symbol_pattern.search(word)) / word_count
""", block="block", language="python"),
),
H5("Fraction of Alphabetic Words"),
Details(
Summary("Implementations from Dolma"),
D_code("""
attrs.fraction_of_words_with_alpha_character = sum(
1 for word in words if any(c.isalpha() for c in word)
) / max(word_count, 1)
""", block="block", language="python"),
),
Details(
Summary("Implementations from RedPajama-V2"),
D_code("""
class RPS_Doc_Frac_No_Alph_Words(RPSBase): # noqa
ALPH_REGEX = re.compile(r"[a-zA-Z]")
__slots__ = ()
def __call__(self, document: Document) -> SignalType:
num_words = document.num_raw_words
if num_words == 0:
return [(0, len(document), None)]
num_words_with_alpha = float(sum(
int(self.ALPH_REGEX.search(word) is not None)
for word in document.raw_words
))
score = 1.0 - num_words_with_alpha / num_words
score = round(score, PRECISION)
return [(0, len(document), score)]
""", block="block", language="python"),
),
Details(
Summary("Implementations from DataTrove"),
D_code("""
# that 80 % of words in a document contain at least one alphabetic character
if (
self.max_non_alpha_words_ratio
and sum([any((c.isalpha() for c in w)) for w in words]) / n_words < self.max_non_alpha_words_ratio
):
return False, "gopher_below_alpha_threshold"
""", block="block", language="python"),
),
P("""
Both Dolma and DataTrove use `char.isalpha()` to detect whether a word contains alphabetic characters while
RedPajama-V2 employs regular expressions for this purpose. We opt to use regular expressions since `char.isalpha()`
can also match words in other languages as long as they are not punctuations.
"""),
H5("Number of Stop Words"),
P("""
The implementations across existing pipelines are largely identical. We adopt them and apply them to our pipeline.
"""),
D_code("""
STOP_WORDS = ('the', 'be', 'to', 'of', 'and', 'that', 'have', 'with')
...
stop_words_pattern = re.compile("|".join(re.escape(symbol) for symbol in STOP_WORDS))
...
attrs.num_of_stop_words = sum(1 for word in words if stop_words_pattern.search(word))
""", block="block", language="python"),
H5("Our Implementations"),
Details(
Summary("Sample documents that are filtered out by statistics-based heuristics"),
DV(
"data/sample_doc_stat.json",
0,
"Sample documents that are filtered out by statistics-based heuristics",
),
),
H4("3.4 Others"),
P("""
Following C4, we remove any page where the phrase “lorem ipsum” appeared since some pages had placeholder “lorem ipsum”
text.
"""),
Details(
Summary("Sample documents containing 'lorem ipsum'"),
DV("data/lorem_ipsum.json", 0, "Sample documents containing 'lorem ipsum'"),
),
H3("4. Deduplication"),
P("""
After careful filtering, although data quality has improved, a large fraction of the content is repeated across documents. This may be due to the crawler indirectly hitting the same page multiple times, to boilerplate content being repeated (e.g., licences), or even to plagiarism. These duplicates can strongly impact models, favoring memorization instead of generalization.
"""), # Add detailed content and images as needed
P("We perform two-level deduplication: local exact deduplication and global fuzzy deduplication"),
P(B("Local Exact Deduplication")),
P("To reduce the expensive cost of global deduplication, we apply a local exact deduplication before it. Specifically, each dump is split into 70 splits. A bloom filter is applied within each split."),
P(B("Global Fuzzy Deduplication")),
P("NEED TO UPDATE"),
H3("5. PII Removal"),
P("..."), # Add detailed content and images as needed
H2("Reference"),
Ul(
Li(
P(
"The {P}ile: An 800{GB} dataset of diverse text for language modeling Gao, Leo and Biderman, Stella and Black, Sid and Golding, Laurence and Hoppe, Travis and Foster, Charles and Phang, Jason and He, Horace and Thite, Anish and Nabeshima, Noa and others. 2020."
)
),
Li(
P("""Scaling Language Models: Methods, Analysis & Insights from Training Gopher [link]
Jack W. Rae and Sebastian Borgeaud and Trevor Cai and Katie Millican and Jordan Hoffmann and H. Francis
Song and John Aslanides and Sarah Henderson and Roman Ring and Susannah Young and Eliza Rutherford and Tom
Hennigan and Jacob Menick and Niklas Muennighoff and Aakanksha Naik and Crystal Nam and Matthew E. Peters
and Abhilasha Ravichander and Kyle Richardson and Zejiang Shen and Emma Strubell and Nishant Subramani
and Oyvind Tafjord and Pete Walsh and Luke Zettlemoyer and Noah A. Smith and Hannaneh Hajishirzi and Iz Beltagy
and Dirk Groeneveld and Jesse Dodge and Kyle Lo. 2021.""")
),
Li(
P("""The RefinedWeb Dataset for Falcon LLM: Outperforming Curated Corpora with Web Data, and Web Data Only
Guilherme Penedo and Quentin Malartic and Daniel Hesslow and Ruxandra Cojocaru and Alessandro Cappelli and
Hamza Alobeidli and Baptiste Pannier and Ebtesam Almazrouei and Julien Launay. 2023.""")
),
Li(
P("""🍷 FineWeb: decanting the web for the finest text data at scale [link]
Guilherme Penedo, Hynek Kydlíček, Loubna Ben Allal, Anton Lozhkov, Colin Raffel, Leandro Werra and Thomas Wolf. 2024.""")
),
Li(
P("""Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer
Colin Raffel and Noam Shazeer and Adam Roberts and Katherine Lee and Sharan Narang and Michael Matena and Yanqi Zhou and
Wei Li and Peter J. Liu. 2023.""")
),
Li(
P("""Dolma: an Open Corpus of Three Trillion Tokens for Language Model Pretraining Research
Luca Soldaini and Rodney Kinney and Akshita Bhagia and Dustin Schwenk and David Atkinson and Russell Authur and
Ben Bogin and Khyathi Chandu and Jennifer Dumas and Yanai Elazar and Valentin Hofmann and Ananya Harsh Jha and
Sachin Kumar and Li Lucy and Xinxi Lyu and Nathan Lambert and Ian Magnusson and Jacob Morrison and Niklas Muennighoff and
Aakanksha Naik and Crystal Nam and Matthew E. Peters and Abhilasha Ravichander and Kyle Richardson and Zejiang Shen
and Emma Strubell and Nishant Subramani and Oyvind Tafjord and Pete Walsh and Luke Zettlemoyer and Noah A. Smith and
Hannaneh Hajishirzi and Iz Beltagy and Dirk Groeneveld and Jesse Dodge and Kyle Lo. 2024.""")
),
Li(
P("""RedPajama-Data-v2: an Open Dataset with 30 Trillion Tokens for Training Large Language Models [link]
Together Computer. 2023.""")
),
),
)
|