Spaces:
Sleeping
Sleeping
File size: 14,521 Bytes
1c7fe13 e759b31 230ca5c e759b31 230ca5c e759b31 230ca5c e759b31 230ca5c e759b31 230ca5c e759b31 230ca5c e759b31 230ca5c e759b31 230ca5c e759b31 230ca5c e759b31 230ca5c e759b31 1c7fe13 1703b06 38049d1 0c45d35 9054008 0c45d35 ffdb8be 2acbfea 0c45d35 fcc16a1 0c45d35 0f2f612 0c45d35 6afc890 e759b31 17b2190 1c7fe13 |
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 |
from fasthtml.common import *
from fasthtml.components import *
from fasthtml.components import D_title, D_article, D_front_matter, D_contents, D_byline
from plotly import graph_objects as go
from fh_plotly import plotly2fasthtml
import pandas as pd
import json
from rich import print
import curated
import web
import common
import results
dataset_comparison = 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",
],
"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",
],
}
)
table_html = dataset_comparison.to_html(index=False, border=0)
table_div = Div(NotStr(table_html), 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",
],
}
)
table_html = dataset_sources.to_html(index=False, border=0)
table_div1 = Div(NotStr(table_html), style="margin: 40px;")
def get_curated_chart():
# Dataset
data = {
'Source': ['ArXiv', 'PubMed Central', 'PubMed Abstract', 'S2ORC Full Text', 'S2ORC Abstract', 'PhilPapers', 'Wikipedia', 'StackExchange', 'EuroParl', 'Ubuntu IRC', 'Freelaw', 'PG19', 'USPTO', 'HackerNews', 'DM Maths'],
'Category': ['Papers', 'Papers', 'Papers', 'Papers', 'Papers', 'Papers', 'Internet', 'Conversational', 'Legal/Formal', 'Conversational', 'Legal/Formal', 'Books', 'Legal/Formal', 'Conversational', 'Reasoning'],
'Count': [100, 200, 150, 120, 80, 90, 300, 250, 180, 150, 150, 250, 180, 120, 90],
'Details': [
'A repository of scientific papers in various disciplines, including computer science, physics, mathematics, and more.',
'A database of biomedical and life sciences research articles.',
'Abstracts of biomedical literature from various sources.',
'Full-text articles from the Semantic Scholar Open Research Corpus.',
'Abstracts of articles from the Semantic Scholar Open Research Corpus.',
'Papers from the PhilPapers database, a comprehensive index and bibliography of philosophy research.',
'A collaborative online encyclopedia that covers a wide range of topics.',
'A network of question-and-answer websites on various subjects, including programming, science, mathematics, and more.',
'A collection of multilingual parallel corpora of parliamentary debates from the European Parliament.',
'Chat logs from the Ubuntu Internet Relay Chat (IRC) channels.',
'Legal documents and court cases from various jurisdictions.',
'A collection of books from Project Gutenberg, a digital library of public domain works.',
'Patent documents from the United States Patent and Trademark Office.',
'User-generated news and discussion platform focused on technology and startups.',
'Deep Mind Maths dataset with generated questions.'
]
}
# Calculate percentage for each data source
total_count = sum(data['Count'])
data['Percentage'] = [count / total_count * 100 for count in data['Count']]
# Create treemap
fig = px.treemap(data, path=['Category', 'Source'], values='Count', hover_data=['Details', 'Percentage'], hover_name='Source')
# Set the size of the chart
fig.update_layout(width=800, height=600)
# Display treemap
st.plotly_chart(fig)
overview_div = Div(
Section(
H2("Combining the Best of Web and Curated Sources"),
P("""The quality and size of a pre-training dataset play a crucial role in the performance of large language models (LLMs).
The community has introduced a variety of datasets for this purpose, including purely web-based datasets like RefinedWeb{citation_obj.display_citation("refinedweb")}, RedPajama-Data-V2{citation_obj.display_citation("redpajama-v2")}, DCLM{citation_obj.display_citation("dclm")}, and FineWeb{citation_obj.display_citation("fineweb")},
as well as comprehensive datasets derived from multiple highly-curated data sources such as The Pile{citation_obj.display_citation("thepile")}, RedPajama-Data-V1{citation_obj.display_citation("redpajama-v1")}, and Dolma{citation_obj.display_citation("dolma")}.
It is commonly known that web-based datasets provide a vast quantity of data, while highly-curated multi-source datasets consistently deliver high quality and diversity,
both critical for effective LLM pre-training."""),
P("However, despite the advancements in both types of data, each type of dataset has its limitations. For instance, the processing scripts for the web dataset, RefinedWeb, known for its high quality, are not public, and only about 10% of the entire dataset has been disclosed. Conversely, the web component of existing highly-curated multi-source datasets is relatively small compared to purely web-based datasets, limiting their coverage and diversity compared to the scale of information from the internet."),
P("By integrating the extensive reach of web data with the exceptional quality of curated sources, TxT360 is crafted to meet and surpass the rigorous standards required for state-of-the-art LLM pre-training."),
H3("TxT360 combines both the web data and highly-curated sources, which none of the existing datasets have covered."),
P("Table 1: The following table shows TxT360 and other well-known datasets on the coverage and size of data sources."),
table_div,
P("Table 2: Basic TxT360 Statistics."),
table_div1,
),
Section(
H2("Our General Appoach to Data Processing"),
P("We enforce a fully transparent data processing pipeline when producing TxT360, designed to handle both web and curated datasets with precision and clarity. This transparent pipeline presents a unified framework for processing both data types, making it convenient and adaptive for users to revise and fine-tune the pipeline. "),
P("For web datasets, the pipeline focuses on extracting meaningful, high-quality text from raw web content, which is inherently noisy and varied. The process includes sophisticated filtering and deduplication techniques to clean the data and remove any redundancies or irrelevant information. On the other hand, curated datasets, which are already more structured and reliable, are processed with selective steps to maintain their integrity while integrating them seamlessly 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.")
),
Section(
H2("Global Deduplication"),
P("Deduplication is crucial in language model (LM) pre-training for several reasons. It reduces the volume of training data, leading to shorter training times and potentially better accuracy. It also prevents train-test overlap, improves evaluation metrics, and minimizes the risk of memorization, which can cause test loss to increase during training. By controlling the pretraining data distribution through deduplication and selective upsampling, we avoid relying on the often inconsistent distribution of internet-sourced data."),
P("Our deduplication process began with 61.8 TB of high-quality, 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.")),
),
)
def overview():
return Div(Section(
overview_div,
id="inner-text",
)
)
|