]",
language="bash",
),
". The markdown files are combined to create jsonl files. PubMed Abstract (PMA) files were downloaded in xml. The BeautifulSoup library was used to extract the abstract, title, and PMID. All files were stored in jsonl format.",
),
P(B("Unique Data Preparation Challenges: ")),
Ul(
Li(
"We tried similar attempts on PMC as we did on ArXiv. The resulted markdown might have slight difference due to the different structure of the XML files.",
style="margin-bottom: -3px",
),
),
P(
B("Filters Applied: "),
"Multiple filters are used here after manually verifying output of all the filters as suggested by peS2o dataset.",
),
Ul(
Li(
"Minimum Word Count Filter: PMC documents with less than 100 words (not inclusive) are discarded; PMA documents less than 20 words are discarded",
style="margin-bottom: -3px",
),
Li("Language Filter: English only", style="margin-bottom: -3px"),
Li(
"Frequency Filter: The most frequent word in the paper consists of alpha characters only, and it appears in less than 7.5% of the document. Words are obtained by splitting the text on whitespace. This filter was not used for PMA",
style="margin-bottom: -3px",
),
Li(
"Unigram Log Probability Threshold: -20",
style="margin-bottom: -3px",
),
),
table_div_med,
Details(
Summary("PubMed Filtering Examples"),
Div(
pubmed_examples,
style="background-color: white; padding: 15px; margin-top: 10px; margin-bottom: 10px; border-radius: 8px; border: none; ", # Styling for the DV2 part
),
style="""
background-color: #FFFAEA; /* Light yellow background */
padding: 15px;
border-radius: 12px;
margin-bottom: 15px
""",
),
),
),
Section(
Div(
H3("Phil Papers"),
P(
"Papers from the PhilPapers database, a comprehensive index and bibliography of philosophy research maintained by the Center for Digital Philosophy at the University of Western Ontario."
),
P(
B("Download and Extraction: "),
"Original PDF files download from ",
A(
"https://philarchive.org/oai.pl",
href="https://philarchive.org/oai.pl",
),
". All available PDF's were downloaded. Each PDF was converted to text using java",
D_code(
"-jar ../philpapers_resources/src/pdfbox-app-2.0.21.jar ExtractText {f0} {FOUT.name}",
language="python",
),
". After converting to text formatting, a language was detected and added using the langdetect (citation needed) library.",
),
P(B("Filters Applied: ")),
Ul(
Li(
P(
"Hyphenation Removal:",
D_code("end-of", language="python"),
" becomes ",
D_code("end of", language="python"),
),
style="margin-bottom: -3px",
),
Li(
P(
"Newline Filtering:",
D_code("This is/na sentence.", language="python"),
" becomes ",
D_code("This is a sentence.", language="python"),
),
style="margin-bottom: -3px",
),
Li(
P(
"Header/Footer Filtering:",
D_code("(c) 2023 Company Name.", language="python"),
" is removed ",
),
style="margin-bottom: -3px",
),
Li(
P(
"Double Whitespace Filtering:",
D_code("This is a test.", language="python"),
" becomes ",
D_code("This is a test.", language="python"),
),
style="margin-bottom: -3px",
),
Li(
P(
"Mean Line Length Check: ",
"removes paragraphs with an average line length of < 2.0",
),
style="margin-bottom: -3px",
),
Li(
P(
"CID Percentage Filter: ",
"removes LaTex heavy paragraphs that contain over 10% “CID” font artifacts.",
),
style="margin-bottom: -3px",
),
Li(
P(
"Letterness Filter: ",
"discards paragraphs with a low proportion of letters",
),
style="margin-bottom: -3px",
),
Li(
P(
"Removing Leading/Trailing Numbers: ",
"removes numbers at the start or end of paragraphs. ",
D_code("1 This is a sentence.", language="python"),
" becomes ",
D_code("This is a sentence.", language="python"),
),
style="margin-bottom: -3px",
),
Li(
P("Fixing Unicode Issues: ", "fixes Unicode issues."),
style="margin-bottom: -3px",
),
Li(
P(
"Combining Diacritics Correction: ",
D_code("a'", language="python"),
" becomes ",
D_code("å", language="python"),
),
style="margin-bottom: -3px",
),
Li(
P(
"Unigram Log Probability: ",
"the document must have higher than -20 average unigram log probability.",
),
style="margin-bottom: -3px",
),
),
table_div_phil,
Details(
Summary("Phil Papers Filtering Examples"),
Div(
phil_examples,
style="background-color: white; padding: 15px; margin-top: 10px; margin-bottom: 10px; border-radius: 8px; border: none; ", # Styling for the DV2 part
),
style="""
background-color: #FFFAEA; /* Light yellow background */
padding: 15px;
border-radius: 12px;
margin-bottom: 15px
""",
),
),
),
Section(
Div(
H3("EuroParl"),
P(
"A collection of multilingual parallel corpora of parliamentary debates from the European Parliament. This is a high-quality legacy dataset earlier used for translation tasks."
),
P(
B("Download and Extraction: "),
"Original dataset was downloaded from ",
A(
"http://www.statmt.org/europarl/v7/europarl.tgz",
href="http://www.statmt.org/europarl/v7/europarl.tgz",
),
". The files were converted to jsonl lines for filtering.",
),
P(
B("Filters Applied: "),
"EuroParl was initially filtered during the download process. Documents with fewer than 200 characters were removed. The documents also contained HTML tags which were removed.",
),
D_code(
"""
Raw single line in data: Hi I am speaker
After tag removal: P Hi I am speaker
We remove everything that starts with ["P", "BRK", "CHAPTER", "/P"]
and only keep tagnae == SPEAKER
because line starting with TEXT TEXT ....... has the relevant text
""",
style="block",
language="python",
),
D_code(
"""
def process_tag(original_tag):
tag = original_tag.strip(">").strip("<")
# Skip empty tags
if not tag:
return None
tagname = tag.split()[0]
# Skip paragraph, break, and chapter tags
if tagname in ["P", "BRK", "CHAPTER", "/P"]:
return None
# For speaker tags, return the name
if tagname == "SPEAKER":
soup = bs4.BeautifulSoup(original_tag, "html.parser")
name = soup.speaker["name"]
return name
# Raise a error here if there is a tag we don't know
raise ValueError(f"Unknown tag {tag}")
""",
style="block",
language="python",
),
table_div_up,
Details(
Summary("EuroParl Filtering Examples"),
Div(
eu_examples,
style="background-color: white; padding: 15px; margin-top: 10px; margin-bottom: 10px; border-radius: 8px; border: none; ", # Styling for the DV2 part
),
style="""
background-color: #FFFAEA; /* Light yellow background */
padding: 15px;
border-radius: 12px;
margin-bottom: 15px
""",
),
),
),
Section(
Div(
H3("HackerNews"),
P(
"A dialog-based dataset where user comments on the links as the head post aggregated by Y Combinator."
),
P(
B("Download and Extraction: "),
"The dataset was downloaded from the HackerNews repo here: ",
A(
"https://hacker-news.firebaseio.com/v0/item/",
href="https://hacker-news.firebaseio.com/v0/item/",
),
". The dataset was parsed using the Story ID. In this dataset each post is a story, and each reply is considered subsequent story. Story IDs were considered between ID 1 to 37500000. The URL for all Story IDs was pinged. If that ID returned an error, the ID was removed. Each request was given a 2 second wait to account for network time.",
),
P(
"The HackerNews dataset contains a vast amount of stories and is known for lively discussions. Due to the number of replies a story may contain, only longest comment thread for each story was sampled past level 3. All stories included the title (1st level) and all direct replies (2nd level)."
),
P(B("Unique Data Preparation Challenges: ")),
Ul(
Li(
"As discussed above, the comment heirarchies required a thoughful approach to extracting meaningful data. ",
style="margin-bottom: -3px",
),
Li(
"In the comment thread heirarchy, relationships had to be assigned to between the comments, sub-comments, and original story ID. ",
style="margin-bottom: -3px",
),
),
P(B("Filters Applied: ")),
Ul(
Li("Language Filter: English", style="margin-bottom: -3px"),
Li("Minimum Word Count Filter: 10", style="margin-bottom: -3px"),
Li(
"Unigram Log Probability Threshold: -20",
style="margin-bottom: -3px",
),
),
table_div_hn,
),
),
Section(
Div(
H3("USPTO"),
P("Patent documents from the United States Patent and Trademark Office."),
P(
B("Download and Extraction: "),
"Data was downloaded and extracted using tags from ",
A(
"https://bulkdata.uspto.gov/data/patent/grant/redbook/fulltext/",
href="https://bulkdata.uspto.gov/data/patent/grant/redbook/fulltext/",
),
". There were three different formats that needed three different functions to download and extract the data based on year:",
I("Pre_2002"),
", ",
I("2002_to_2004"),
" and",
I("post_2004"),
". We used the exact code used in The Pile (citation needed).",
),
P(B("Filters Applied: ")),
Ul(
Li("Language Filter: English", style="margin-bottom: -3px"),
Li("Minimum Word Count Filter: 50", style="margin-bottom: -3px"),
Li("Unigram Log Probability", style="margin-bottom: -3px"),
),
table_div_uspto,
),
),
Section(
Div(
H3("FreeLaw"),
P(
"Legal documents and court cases from various jurisdictions provided by US-registered non-profit firm Free Law Project. We have included data from CourtListener which included millions of legal opinions from federal and state courts."
),
P(
B("Download and Extraction"),
"The dataset was downloaded from: ",
A(
"https://storage.courtlistener.com/bulk-data/",
href="https://storage.courtlistener.com/bulk-data/",
),
". There are 19 CSV files which contain overlapping content. CSV files can contain content in multiple columns requiring a holistic extraction approach. Text was extracted from the following using html2text function. The block below shows how each text type was extracted.",
),
D_code(
"""
("html", html2text), ("html_lawbox", html2text),
("html_columbia", html2text), ("html_anon_2020", html2text),
("html_with_citations", html2text), ("xml_harvard", html2text),
plain_text
""",
language="python",
),
P(
"All content was downloaded leading to high number of documents filtered during local deduplication. Following The Pile, priority was given to plain_text first, followed by the columns in the table in reverse order."
),
P(B("Unique Data Preparation Challenges: ")),
Ul(
Li(
"Consecutive whitespaces and tabs were found. Consecutive Whitespaces and tabes were reduce to one, single whitespace.",
style="margin-bottom: -3px",
),
Li(
"Whitespaces were found between new lines with no addition text. These whitespaces were removed.",
style="margin-bottom: -3px",
),
Li(
"Consecutive new lines were found in some documents without leading to a new paragraph. All consecutive newline to a single new line.",
style="margin-bottom: -3px",
),
Li(
"Converted all single new lines to whitespace. If whitespace was found after a new line with no text, the whitespace was removed. All leading and trailing whitespace was removed.",
style="margin-bottom: -3px",
),
Li("All \f characters were removed.", style="margin-bottom: -3px"),
),
P(B("Filters Applied: ")),
Ul(
Li("Language Filter: English", style="margin-bottom: -3px"),
Li("Minimum Word Count Filter: 50", style="margin-bottom: -3px"),
Li("Unigram Log Probability", style="margin-bottom: -3px"),
),
P(
"Note: Local deduplication within FreeLaw itself removed 90%+ of the dataset as duplicate."
),
table_div_freelaw,
Details(
Summary("FreeLaw Filtering Examples"),
Div(
freelaw_examples,
style="background-color: white; padding: 15px; margin-top: 10px; margin-bottom: 10px; border-radius: 8px; border: none; ", # Styling for the DV2 part
),
style="""
background-color: #FFFAEA; /* Light yellow background */
padding: 15px;
border-radius: 12px;
margin-bottom: 15px
""",
),
),
),
Section(
Div(
H3("StackExchange"),
P(
"A network of question-and-answer websites on various subjects, including programming, science, mathematics, and more. This is one of the largest publicly available repositories for question-answer pairs. We have included comments also to include an overall discussion on each post."
),
P(
B("Download and Extraction: "),
"The archive dataset was used to download all data from StackExchange and 364 StackExchange's sub URLs including: ",
A("math.stackexchange.com", href="math.stackexchange.com"),
". Raw data was extracted an XML format and only two files Posts.xml and Comments.xml were considered. To match the StackExchange hierarchy, each file was parsed using post_id to connect questions to answers and then to comments. We will include the full list of sub URLs in when the code is released.",
),
D_code(
"""
1. Questions:
2. Comment1:
3. Comment2:
4. Answer1:
5. Comment1:
6. Comment2:
7. Answer2:
8. Comment1:
9. Comment2:""",
block="block",
language="python",
),
P(B("Unique Data Preparation Challenges: ")),
Ul(
Li(
"Handling code block was a required finding the specific blocks and exacting the details in one snippet.",
style="margin-bottom: -3px",
),
Li(
"Question and Answer formatting had to be rewritten to match the question and the anwer.",
style="margin-bottom: -3px",
),
Li(
"Occasionally a title was not included at the beginning of a question. For consistent formatting, a title was added.",
style="margin-bottom: -3px",
),
),
P(B("Filters Applied: ")),
Ul(
Li("Minimum Word Count Filter: 10", style="margin-bottom: -3px"),
),
table_div_se,
Details(
Summary("StackExchange Filtering Examples"),
Div(
se_examples,
style="background-color: white; padding: 15px; margin-top: 10px; margin-bottom: 10px; border-radius: 8px; border: none; ", # Styling for the DV2 part
),
style="""
background-color: #FFFAEA; /* Light yellow background */
padding: 15px;
border-radius: 12px;
margin-bottom: 15px
""",
),
),
),
Section(
Div(
H3("Ubuntu IRC"),
P(
"Chat logs from the Ubuntu Internet Relay Chat (IRC) channels on the Freenode IRC chat server. This data is also another form of dialog dataset on niche topics."
),
P(
B("Download and Extraction: "),
"The dataset was downloaded from: ",
A(
"https://irclogs.ubuntu.com/{date.year}/{date.month:02d}/{date.day:02d}/",
href="https://irclogs.ubuntu.com/{date.year}/{date.month:02d}/{date.day:02d}/",
),
" based on the year.",
),
P("During extraction, the logs were cleaned using following functions:"),
D_code(
"""
def exclude_system(x):
return '\n'.join(line for line in x.split('\n') if not line.startswith('==='))
def exclude_select_system(x):
return '\n'.join(line for line in x.split('\n') if not (line.startswith('===')
and any(term in line for term in
['has joined #', 'has left #', 'Topic for #', "Topic (#", "is now known as"]) ))
def clean(x):
return '\n'.join('* ' + line[4:] if line.startswith('===') else line[8:] for line in x.split('\n'))
""",
block="block",
language="python",
),
P(B("Unique Data Preparation Challenges: ")),
Ul(
Li(
"Similar to the HackerNews challenges, we had to map comments and sub-comments to the original question.",
style="margin-bottom: -3px",
),
),
P(B("Filters Applied: ")),
Ul(
Li("Language Filter: English", style="margin-bottom: -3px"),
Li("Minimum Word Count Filter: 10", style="margin-bottom: -3px"),
Li("Unigram Log Probability", style="margin-bottom: -3px"),
),
table_div_uirc,
),
),
Section(
Div(
H3("DM Math"),
P(
"DeepMind Math dataset with generated questions from various topics like algebra, calculus, geometry, etc. Maths data is included to improve model reasoning abilities in the downstream tasks."
),
P(
B("Download and Extraction: "),
"The dataset was downloaded directly from the Huggingface repo: ",
A(
"https://huggingface.co/datasets/deepmind/math_dataset",
href="https://huggingface.co/datasets/deepmind/math_dataset",
),
". The data was converted to the jsonl format where lines is represented as:",
),
D_code(
"""
Question: TEXT
Answer: TEXT""",
block="block",
language="python",
),
P(B("Unique Data Preparation Challenges: ")),
Ul(
Li(
"A byte string was included at the beginning of new lines",
style="margin-bottom: -3px",
),
Li('No space before keyword "Answer:"', style="margin-bottom: -3px"),
),
P(B("Filters Applied: ")),
Ul(
Li("No filtering was applied to DM Math", style="margin-bottom: -3px"),
),
table_div_dmm,
Details(
Summary("DM Math Filtering Examples"),
Div(
dmm_examples,
style="background-color: white; padding: 15px; margin-top: 10px; margin-bottom: 10px; border-radius: 8px; border: none; ", # Styling for the DV2 part
),
style="""
background-color: #FFFAEA; /* Light yellow background */
padding: 15px;
border-radius: 12px;
margin-bottom: 15px
""",
),
),
),
Section(
Div(
H3("PG-19"),
P(
"A collection of books from Project Gutenberg, a digital library of public domain works. This contains all the books that were published before 1919."
),
P(
B("Download and Extraction: "),
"The dataset was downloaded directly from Huggingface: ",
A(
"https://huggingface.co/datasets/deepmind/pg19",
href="https://huggingface.co/datasets/deepmind/pg19",
),
".",
),
P(B("Unique Data Preparation Challenges: ")),
Ul(
Li(
"Consecutive whitespaces were found spanning 10+ whitespace entries. These whitespaces were reduce to one, single whitespace.",
style="margin-bottom: -3px",
),
Li(
"Consecutive new lines were found in some documents. All consecutive news over two were were reduce to two new lines.",
style="margin-bottom: -3px",
),
Li(
"Delimiters such as * * * * * * * * ? were found. They were removed and replaced with whitespace.",
style="margin-bottom: -3px",
),
),
P(B("Filters Applied:")),
Ul(
Li("Language Filter: English", style="margin-bottom: -3px"),
Li("Minimum Word Count Filter: 20", style="margin-bottom: -3px"),
Li("Unigram Log Probability: ", "-20", style="margin-bottom: -3px"),
),
table_div_pg19,
Details(
Summary("PG-19 Filtering Examples"),
Div(
pg19_examples,
style="background-color: white; padding: 15px; margin-top: 10px; margin-bottom: 10px; border-radius: 8px; border: none; ", # Styling for the DV2 part
),
style="""
background-color: #FFFAEA; /* Light yellow background */
padding: 15px;
border-radius: 12px;
margin-bottom: 15px
""",
),
),
),
)
local_dedup_text = P(
"Each curated data source has been prepared using its specific rules and has been locally deduped using min-hash near deduplication. Details about the dataset are shown below in the table:"
)
data_pipeline_table = pd.DataFrame(
{
"Data Source": [
"Papers",
"Wikipedia",
"StackExchange",
"Europarl",
"Ubuntu IRC",
"HackerNews",
"PG-19",
"USPTO",
"Freelaw",
"DM Math",
],
"Percent Filtered": [
"15%",
"21%",
"<0.1%",
"1%",
"0.4%",
"60%",
"0.8%",
"22.5%",
"94%",
"0",
],
"Unique Document Percentage": [
"75.99%",
"91.91%",
"98.02%",
"98.87%",
"100%",
"99.91%",
"31.81%",
"99.94%",
"91.01%",
"0",
],
"2 - 5 Duplicates": [
"19.4%",
"4.7%",
"1.27%",
"0.94%",
"0",
"0.05%",
"20.03%",
"0.05%",
"6,87%",
"0",
],
"6 - 10 Duplicates": [
"2.89%",
"1.58%",
"0.35%",
"0.09%",
"0",
"0.02%",
"24.27%",
"0.01%",
"1.07%",
"0",
],
"11 - 100 Duplicates": [
"1.17%",
"1.76%",
"0.35%",
"0.1",
"0",
"0.02%",
"22.26%",
"0.01%",
"1.05%",
"0",
],
"101 - 1000 Duplicates": [
"0.01%",
"0.05%",
"0.01%",
"0",
"0",
"<0.01%",
"1.58%",
"<0.01%",
"0.01%",
"0",
],
"1001+ Duplicates": [
"<0.01%",
"<0.01%",
"<0.01%",
"0",
"0",
"<0.01%",
"0.06%",
"0",
"0",
"0",
],
}
)
table_html_data_pipe = data_pipeline_table.to_html(index=False, border=0)
table_div_data_pipe = Div(NotStr(table_html_data_pipe), style="margin: 40px;")
def curated():
data_preparation_steps = pd.DataFrame(
{
"Method": [
"HTTP/FTP dumps",
"Web crawling",
"Archive snapshot",
"Generated",
"Curated",
],
"Description": [
"Acquiring data from HTTP/FTP dumps",
"Crawling websites to extract data",
"Working with archive dumps",
"Generating synthetic data",
"High quality curated data",
],
"Source": [
"Freelaw | Wikipedia | PhilPapers | Arxiv | S2ORC | Pubmeds",
"USPTO | Hackernews | Ubuntu IRC",
"StackExchange",
"DM Maths",
"PG19 | Europarl",
],
}
)
table_html = data_preparation_steps.to_html(index=False, border=0)
table_div = Div(NotStr(table_html), style="margin: 40px;")
text = P(
"""This initial stage serves as the foundation for the entire
process. Here, we focus on acquiring and extracting the raw data, which can
come from various sources such as crawling websites, using HTTP/FTP dumps,
or working with archive dumps. For instance, to download and prepare a
dataset, we can specific downloaders based on the data source. Each dataset
might have its own downloader script which can be updated in real time to
handle changes in the data source. Here is a general outline of the data
preparation process: It is worth noting that some pipelines might require
invoking additional functions or scripts to handle specific data sources or
formats. These helper scripts can be located within specific directories
or modules dedicated to the dataset."""
)
return Div(
Section(
overview,
),
Section(
curated_sources_intro,
plotly2fasthtml(treemap_chart),
id="section31",
),
Section(
data_preprocessing_div,
id="section32",
),
Section(
filtering_process,
id="section33",
),
id="inner-text",
)