File size: 15,966 Bytes
e137e27
 
005657d
 
 
 
 
 
 
 
 
 
e137e27
 
 
 
 
8262fca
e137e27
 
 
 
005657d
87a6313
e137e27
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
005657d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
e137e27
 
 
 
 
3d1994e
e137e27
 
 
 
 
 
 
 
 
005657d
e137e27
 
 
 
 
43e1d29
 
 
e5180e3
43e1d29
 
e137e27
 
 
 
 
5d3f993
e137e27
 
 
 
 
 
 
b2b380b
e137e27
 
 
 
 
 
 
b2b380b
e137e27
 
 
 
 
 
 
227158f
fac35b0
 
e5180e3
 
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
3d1994e
fac35b0
 
3d1994e
fac35b0
 
3d1994e
fac35b0
 
3d1994e
fac35b0
8580754
e137e27
5025d3d
 
 
 
 
 
 
3d1994e
 
 
 
 
5025d3d
 
 
 
 
 
 
 
3d1994e
5025d3d
 
3d1994e
5025d3d
 
3d1994e
5025d3d
3d1994e
 
 
 
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
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: a globally deduplicated dataset for LLM pretraining",
                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#section1",
                            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",
                                )
                            ),
                        ),
                    ),
                    Div(
                        A(
                            "Global Processing Steps",
                            href="/common#section1",
                            hx_get="/common#section1",
                            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 TxT360 (Trillion eXtracted Text) the first dataset to globally deduplicate 99 CommonCrawl snapshots and 14 commonly used non-web data sources (e.g. FreeLaw, PG-19, etc.) providing pretraining teams with a recipe to easily adjust data weighting and train the most performant models."
            ),
            P(
                "Building on top of the prior studies on pre-training data, TxT360 carefully implements data processing steps including extraction, filtering, deduplication, personally identifiable information removal, and other steps."
            ),
            P(
                "Metadata is stored to recover the raw distribution for each dataset, enabling 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 documented all implementation details in this blog post and are open sourcing the code. Examples of each filter and rationale supporting each decision are included."
            ),
            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."
            ),
            table_div_1,
            table_div_2,
            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. Combining both datasets plays a critical role for effective LLM pre-training. 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. See Results section below."
            ),
            #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. 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. 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 5.7T tokens of high-quality data."
            ),
            P(
                "We provide details and context for the choices behind TxT360 in the respective Web Data Processing and Curated Source Processing section. A deep dive in the deduplication [here]. "
            ),
            #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",
        ),
        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()