Spaces:
Running
Running
File size: 22,396 Bytes
e137e27 005657d e137e27 8262fca e137e27 005657d 87a6313 e137e27 005657d e137e27 591cd18 e137e27 005657d e137e27 43e1d29 e137e27 5d3f993 e137e27 b2b380b e137e27 b2b380b e137e27 b2b380b e137e27 227158f fac35b0 43e1d29 fac35b0 227158f e137e27 fac35b0 43e1d29 fac35b0 e137e27 fac35b0 43e1d29 fac35b0 e137e27 fac35b0 43e1d29 fac35b0 e137e27 005657d e137e27 8580754 fac35b0 8580754 12ce41f fac35b0 b6d74c9 140edc3 8580754 fac35b0 12ce41f fac35b0 b6d74c9 a1001c2 140edc3 8580754 fac35b0 12ce41f a1001c2 fac35b0 0e10a03 a1001c2 140edc3 8580754 fac35b0 8580754 e137e27 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 |
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
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 = """
<d-front-matter>
<script id='distill-front-matter' type="text/json">{
"title": "",
"description": "",
"published": "",
"affiliation": {},
"authors": [
{
"author":"",
"authorURL":""
}
],
"katex": {
"delimiters": [
{"left": "$$", "right": "$$", "display": false}
]
}
}
</script>
</d-front-matter>
"""
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():
return Div(
D_title(
H1(
"TxT360: the most comprehensive, highest quality, and production ready pretraining dataset",
cls="l-body",
style="text-align: center;",
),
Div(
Img(src="images/llm360_logo.png"),
id="title-plot",
cls="main-plot-container l-page",
),
),
Div(D_byline(), NotStr(front_matter), style="display: none;"),
D_article(
D_contents(
Nav(
H3("Table of Contents"),
Div(
A(
"TxT360",
href="/intro#section1",
hx_get="/intro",
hx_target="#inner-text",
)
),
Div(
Ul(
Li(
A(
"About TxT360",
href="/intro#section1",
hx_get="/intro#section1",
hx_target="#inner-text",
)
),
Li(
A(
"Motivation Behind Txt360",
href="/intro#section2",
hx_get="/intro#section2",
hx_target="#inner-text",
)
),
Li(
A(
"Generalizable Approach to Data Processing",
href="/intro#section3",
hx_get="/intro#section3",
hx_target="#inner-text",
)
),
Li(
A(
"Introducing Global Deduplication",
href="/intro#section4",
hx_get="/intro#section4",
hx_target="#inner-text",
)
),
),
),
Div(
A(
"Global Processing Steps",
href="/common",
hx_get="/common",
hx_target="#inner-text",
)
),
Div(
A(
"Web Data Processing",
href="/webdata",
hx_get="/webdata",
hx_target="#inner-text",
)
),
Div(
A(
"Curated Sources Processing",
href="/curated",
hx_get="/curated",
hx_target="#inner-text",
)
),
Div(
A(
"TxT360 Results",
href="/results",
hx_get="/results",
hx_target="#inner-text",
),
),
role="navigation",
cls="l-text figcaption",
),
),
intro(),
),
D_appendix(D_bibliography(src="bibliography.bib")),
Div(*read_bibs(), style="display: none;"),
)
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"
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,
)
.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,
)
.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;")
dataset_sources = pd.DataFrame(
{
"Data Source": [
"CommonCrawl",
"Papers",
"Wikipedia",
"Freelaw",
"DM Math",
"USPTO",
"PG-19",
"HackerNews",
"Ubuntu IRC",
"Europarl",
"StackExchange",
],
"Raw Data Size": [
"11 TB",
"712 GB",
"210 GB",
"23 GB",
"22 GB",
"45 GB",
"11 GB",
"4.1 GB",
"4.7 GB",
"6.1 GB",
"45 GB",
],
"Token Count": [
"5.71T",
"154.96B",
"4.75B",
"7.34B",
"5.23B",
"4.95B",
"2.94B",
"1.08B",
"1.54B",
"1.96B",
"8.37B",
],
"Cut-Off Date": [
"2024-30",
"Q4 2023",
"-",
"Q1 2024",
"-",
"Q4 2023",
"-",
"Q4 2023",
"Q4 2023",
"-",
"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,
).hide(axis="index") # Hide the row index
table_html_data = styled_table._repr_html_()
# table_html_data = dataset_sources.to_html(index=False, border=0)
table_div_data = Div(NotStr(table_html_data), style="margin: 40px;")
@app.get("/intro")
def intro():
return Div(
Section(
H2("About TxT360"),
P(
"We introduce Trillion eXtracted Text (TxT360): the highest quality, most comprehensive, and production ready pretraining dataset released to date, an open source pre training textual dataset with the widest coverage to date, including all 99 Common Crawl snapshots released to date and 14 widely used curated sources such as Wikipedia, StackExchange, and ArXiv."
),
P(
"Building on top of the prior study on pre-training data, (cite C4, chinchilla, fineweb, refinedweb, 2305.16264, ) TxT360 carefully implements data processing steps such as extraction, filtering, deduplication, PII removal and so on. Notably, we design a schema to provide users with full control on the weights of each data point, instead of relying on unknown distribution from the raw dataset. Specifically, we strive to make each data point appear only once, via global deduplication across the whole corpus. Importantly, we store enough metadata to recover the raw distribution, and additionally allow one to have fine-grained control to create data distributions and corpus of desired size. As an example, we present one simple upsampling scheme that takes into account the duplication counts, resulting in a 15~16 trillion token corpus, outperforming FineWeb and our non-upsampling baselines, on diverse evaluations. Unlike DCLM and RedPajama V2, we present the final deduplicated dataset that is ready to go."
),
P(
"We have carefully implemented a large set of filters by carefully reviewing the input, output from data samples and tailored design for each data source. We learned from and compared our implementation with prior data pipelines, such as Chinchilla, C4, DataTrove, Dolma, and RedPajama Data V2 to make carefully considered decisions."
),
P(
"In light of the 360 open source spirit, we will open source the code and document all the implementation details in this blog post, along with the examples and why each decision is made, hoping to shed some light on this important but tedious step."
),
id="section1",
),
Section(
H2("Motivation Behind Txt360"),
H3(
"TxT360 is the first dataset to combine both web and curated data sources commonly used in pretraining."
),
P(
"The quality and size of a pre-training dataset play a crucial role in the performance of large language models (LLMs). Data is often referred as low quality if it has not been filtered to review unwanted text. The community has introduced a variety of filtered datasets including purely web-based datasets."
),
P(
"In pretraining, it is common to combine web data and curated sources (cite). Web data is included to provide a vast quantity of long tail and diverse data, while curated datasets are often information rich and provide the 'deep-dive' domain information. Both datasets play critical for effective LLM pre-training."
),
H4("The Gap TxT360 Fills"),
P(
"Despite advancements in filtering and source material for both data types, each type of dataset has its limitations. RefinedWeb is known for its high quality content but and only about 10% of the entire dataset has been disclosed and the processing scripts have not been released. For datasets that have combined curated sources with web data, the web component is relatively small (NEED TO UPDATE - citation needed)."
),
P(
"By integrating the extensive reach of web data with the exceptional quality of curated sources, TxT360 meets and surpasess the rigorous standards required for state-of-the-art LLM pre-training as demostated in the Results section."
),
P(
"Table 1: The following table shows TxT360 and other well-known datasets on the coverage and size of data sources."
),
table_div_1,
table_div_2,
P("Table 2: Basic TxT360 Statistics."),
table_div_data,
id="section2",
),
Section(
H2("Our Generalizable Approach to Data Processing"),
P(
"To produce TxT360, a comprehensive and transparent 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. Thus, the pipeline includes sophisticated filtering and deduplication techniques to clean the data and remove redundancies or irrelevant information."
),
P(
"Curated datasets are typically already structured and consistently formatted. TxT360 filters these sources with selective steps to maintain their integrity while providing seamless integration into the larger dataset."
),
P(
"We will open-source the scripts for the whole pipeline, allowing the community to review, replicate, and build upon our processes."
),
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="section3",
),
Section(
H2("Introducing Global Deduplication"),
P(
"Deduplication is crucial in language model pre-training for several reasons (NEEDS CITATION). Main hypothesis on why deduplicaiton is necessary and help include:"
),
Ul(
Li(
"reducing data volume reduces training time (and cost)",
style="margin-bottom: 5px",
),
Li(
"removing duplicataive data can lead to better accuracy",
style="margin-bottom: 5px",
),
Li("prevent train-test overlap", style="margin-bottom: 5px"),
Li(
"minimizes the risk of memorization leading to test loss",
style="margin-bottom: 5px",
),
),
P(
"Furthermore, by controlling the pretraining data distribution through deduplication and selective upsampling avoida relying on the often inconsistent distribution of internet-sourced data."
),
H3("Highlights of the TxT360 Deduplication Process"),
P(
"Our deduplication process began with 61.8 TB of filtered, and compressed documents, totaling approximately 48.83 billion documents. We first performed exact deduplication using a Bloom filter, reducing the dataset by 17% to 40.21 billion documents. For global near-deduplication, we scaled methodologies from prior works like SlimPajama to handle the entire dataset, including 87 Common Crawl dumps and other curated data. This involved generating document signatures, matching them to identify near-duplicates, and clustering these to retain only one document per cluster. We prioritized retaining documents from curated sources and more recent CommonCrawl dumps. Throughout the deduplication stages, we extensively used Dask for distributed data processing."
),
P(
"After the global near-deduplication of all 87 CommonCrawl dumps and other curated data, we removed around 85% of the total documents. This leaves us with approximately 4.24 trillion deduplicated tokens, which aligns with what FineWeb has reported for their iterative global deduplication. Along with the list of duplicated documents to delete, our deduplication code also saves some metadata about the duplicate clusters that we find. We save statistics about every duplicate cluster we find, with the document ID of the document we retain from the cluster as the key and with a value capturing the distribution of the duplicates within the cluster over the CommonCrawl dumps (identified by the first 2 digits of every document ID). This way, we always have information about the duplicates we have deleted, allowing us to upsample any data distribution we want for training."
),
P(
"During deduplication, it is not feasible to store all the duplicate clusters we form, but we do save some samples at every size. Here are some observations we made by examining these sample duplicate clusters:"
),
Ul(
Li(
"Smaller components tend to have more overlap in their MinHash bands. The smallest components, which are essentially pairs, consist of exact duplicate documents that local exact deduplication missed."
),
Li(
"When clusters contain three or more documents, incremental changes in the text become apparent. For example, there may be a growing list of personnel over the years."
),
Li(
"In sizable clusters comprising 1,000 or more documents, we observe a trend towards templatization. This involves the recurrent use of standardized language to convey general topics such as terms and conditions, warnings, and disclaimers. Such language is prevalent on commercial websites, offering a consistent and efficient way to communicate commonly encountered information."
),
),
P(
"All Deduplication details can be found in the Global Data Processing Steps section."
),
id="section4",
),
id="inner-text",
)
rt("/update/{target}")(data_viewer.update)
rt("/curated")(curated.curated)
rt("/webdata")(web.web_data)
rt("/common")(common.common_steps)
rt("/results")(results.results)
serve()
|