curated.py

from fasthtml.common import *
from fasthtml.components import *
from plotly import graph_objects as go
from fh_plotly import plotly2fasthtml
import pandas as pd
import json
from data_viewer import view_data, gen_random_id
from rich import print
import uuid

overview_text = P("Curated sources comprise high-quality datasets that contain domain-specificity. These sources, such as Arxiv, Wikipedia, and Stack Exchange, provide valuable data that is excluded from the web dataset mentioned above. Analyzing and processing non-web data can yield insights and opportunities for various applications. Details about each of the sources are provided below. ")
copyright_disclaimer = P("We respect the copyright of the data sources and have not included the controversial data that was used in Pile like YouTube and Opensubtitles, Reddit threads, and books.")


dedup_text = 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.")
dedup_text1 = 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.")
dedup_text2 = 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.")          
dedup_text3 = 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:")

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;")

data_descriptions = pd.DataFrame(
    {
        "Source": [
            "Papers - ArXiv",
            "Papers - PhilPapers",
            "Papers - S2ORC",
            "Papers - PubMed Central",
            "Papers - PubMed Abstract",
            "Wikipedia",
            "StackExchange",
            "EuroParl",
            "Ubuntu IRC",
            "Freelaw",
            "PG-19",
            "USPTO",
            "HackerNews",
            "DM Maths",
        ],
        "Description": [
            "The ArXiv dataset is a vast collection of preprint research papers primarily in Mathematics, Computer Science, and Physics. Established in 1991, it offers high-quality text and mathematical knowledge, making it an invaluable resource for academic and scientific research. ArXiv papers are typically written in LaTeX, a popular typesetting system for these fields. We have extracted the information from latex and converted it into a text format.",
            "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.",
            "The Semantic Scholar Open Research Corpus (S2ORC) is a comprehensive dataset designed for natural language processing (NLP) and text-mining research over scientific papers. It includes rich metadata, and abstract and full-text content for millions of academic papers across various disciplines. This dataset is further divided into two components, S2ORC abstract and S2ORC full text.",
            "The PubMed Central (PMC) dataset is a comprehensive collection of full-text biomedical and life sciences journal articles run by the United States of America’s National Center for Biotechnology Information (NCBI). It provides open access to a wealth of scientific literature, facilitating research and discovery in the medical and biological fields starting from 2008 by the NIH Public Access Policy. Articles in PMC are available for text mining and other secondary analyses, making it an invaluable resource for researchers and developers and other downstream tasks.",
            "Abstracts of more than 30 million publications of biomedical literature from various sources mainly including biomedical articles run by the National Library of Medicine. ",
            "Wikipedia is an encyclopedia form of high-quality text data used for language modeling. We have included filtered and deduplicated versions of complete Wikipedia data directly provided by the Wikipedia Foundation for more than 350 languages.",
            "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.",
            "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.",
            "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.",
            "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.",
            "A collection of books from Project Gutenberg, a digital library of public domain works. This contains all the books that were published before 1919.",
            "Patent documents from the United States Patent and Trademark Office.",
            "High-quality dialog-based dataset where user comments on the links as the head post aggregated by Y Combinator.",
            "DeepMind Maths 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.",
        ],

    }
)

table_html_desc = data_descriptions.to_html(index=False, border=0)
table_desc = Div(NotStr(table_html_desc), style="margin: 40px;")


data_sources = [
    "Freelaw",
    "Wikipedia",
    "PhilPapers",
    "Arxiv",
    "S2ORC",
    "S2ORC Abstract",
    "Pubmed",
    "USPTO",
    "Hackernews",
    "Ubuntu IRC",
    "StackExchange",
    "DM Maths",
    "PG19",
    "Europarl",
]


def get_data(data_source: str = "Freelaw", doc_id: int = 3, target: str = "foo"):
    doc_id = max(0, min(int(doc_id), 9))

    if data_source == "Freelaw":
        raw_sample_doc = json.load(open("data/curated_samples/freelaw_raw.json"))
        extracted_sample_doc = json.load(
            open("data/curated_samples/freelaw_extract.json")
        )
    elif data_source == "Wikipedia":
        raw_sample_doc = extracted_sample_doc = json.load(
            open("data/curated_samples/wiki.json")
        )
    elif data_source == "StackExchange":
        raw_sample_doc = json.load(open("data/curated_samples/stackexchange_raw.json"))
        extracted_sample_doc = json.load(
            open("data/curated_samples/stackexchange_extract.json")
        )
    elif data_source == "PhilPapers":
        raw_sample_doc = extracted_sample_doc = json.load(
            open("data/curated_samples/philpapers_raw.json")
        )
    elif data_source == "Arxiv":
        raw_sample_doc = json.load(open("data/curated_samples/arxiv_raw.json"))
        extracted_sample_doc = json.load(
            open("data/curated_samples/arxiv_extract.json")
        )
    elif data_source == "S2ORC":
        raw_sample_doc = extracted_sample_doc = json.load(
            open("data/curated_samples/s2orc_raw.json")
        )
    elif data_source == "S2ORC Abstract":
        raw_sample_doc = extracted_sample_doc = json.load(
            open("data/curated_samples/s2orc_abstract_raw.json")
        )
    elif data_source == "Pubmed":
        raw_sample_doc = json.load(open("data/curated_samples/pubmed_raw.json"))
        extracted_sample_doc = json.load(
            open("data/curated_samples/pubmed_extract.json")
        )
    elif data_source == "DM Maths":
        raw_sample_doc = json.load(open("data/curated_samples/dm_maths_raw.json"))
        extracted_sample_doc = json.load(
            open("data/curated_samples/dm_maths_extract.json")
        )
    elif data_source == "PG19":
        raw_sample_doc = extracted_sample_doc = json.load(
            open("data/curated_samples/pg19_raw.json")
        )
    elif data_source == "Europarl":
        raw_sample_doc = extracted_sample_doc = json.load(
            open("data/curated_samples/europarl_raw.json")
        )
    else:
        raw_sample_doc = extracted_sample_doc = [{} for _ in range(10)]

    raw_json = raw_sample_doc[doc_id]
    extracted_json = extracted_sample_doc[doc_id]
    return view_data(
        raw_json,
        extracted_json,
        doc_id=doc_id,
        data_source=data_source,
        data_sources=data_sources,
        target=target,
    )


def get_chart_28168342():
    fig = go.Figure()
    filter_names = [
        "Download",
        "Language",
        "Min word count",
        "Title Abstract",
        "Majority language",
        "Paragraph count",
        "Frequency",
        "Unigram log probability",
        "Local dedup",
    ]

    data_sources = [
        ("Wikipedia", [100, 90, 80, 70, 60, 50, 40, 30, 20]),
        ("Freelaw", [100, 90, 80, 70, 60, 50, 40, 20, 20]),
        ("DM Maths", [100, 90, 80, 70, 60, 40, 40, 30, 20]),
        ("USPTO", [100, 90, 80, 70, 60, 40, 40, 30, 20]),
        ("PG19", [100, 90, 80, 70, 60, 40, 40, 30, 20]),
        ("Hackernews", [100, 90, 80, 70, 60, 40, 40, 30, 20]),
        ("Ubuntu IRC", [100, 90, 80, 70, 60, 40, 40, 30, 20]),
        ("Europarl", [100, 90, 80, 70, 60, 40, 40, 30, 20]),
        ("StackExchange", [100, 90, 80, 70, 60, 40, 40, 30, 20]),
        ("Arxiv", [100, 90, 80, 70, 60, 40, 40, 30, 20]),
        ("S2ORC", [100, 90, 80, 70, 60, 40, 40, 30, 20]),
        ("S2ORC Abstract", [100, 90, 80, 70, 60, 40, 40, 30, 20]),
        ("PubMed Central", [100, 90, 80, 70, 60, 40, 40, 30, 20]),
        ("PubMed Central Abstract", [100, 90, 80, 70, 60, 40, 40, 30, 20]),
        ("PhilPapers", [100, 90, 80, 70, 60, 40, 40, 30, 20]),
    ]

    for name, x_values in data_sources:
        fig.add_trace(
            go.Funnel(
                name=name,
                orientation="h",
                y=filter_names,
                x=x_values,
                textinfo="value+percent total",
                textposition="inside",
            )
        )

    fig.update_layout(height=500, plot_bgcolor="rgba(0,0,0,0)")
    return fig


def curated(request):
    # Partial Updates
    params = dict(request.query_params)
    if target := params.get("target"):
        if data_source := params.get(f"data_source_{target}"):
            return get_data(
                data_source, params.get(f"doc_id_{target}", 3), params.get("target")
            )
        if doc_id := params.get(f"doc_id_{target}"):
            return get_data(
                params.get(f"data_source_{target}"), doc_id, params.get("target")
            )

    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's 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.""")

    data_preparation_div = Div(
        H3("Data Preparation"),
        text,
        table_div,
        Div(
            get_data(target=gen_random_id()),
            style="border: 1px solid #ccc; padding: 20px;",
        ),
    )

    text = P("""Data preprocessing is a crucial step in the data science
    pipeline. It involves cleaning and transforming raw data into a format that
    is suitable for analysis. This process includes handling missing values,
    normalizing data, encoding categorical variables, and more.""")

    preprocessing_steps = pd.DataFrame(
        {
            "Step": [
                "Language Filter",
                "Min Word Count",
                "Title Abstract",
                "Majority Language",
                "Paragraph Count",
                "Frequency",
                "Unigram Log Probability",
            ],
            "Description": [
                "Filtering data based on language",
                "Setting a minimum word count threshold",
                "Extracting information from the title and abstract",
                "Identifying the majority language in the dataset",
                "Counting the number of paragraphs in each document",
                "Calculating the frequency of each word in the dataset",
                "Calculating the log probability of each unigram",
            ],
            "Need": [
                "To remove documents in unwanted languages",
                "To filter out documents with very few words",
                "To extract relevant information for analysis",
                "To understand the distribution of languages in the dataset",
                "To analyze the structure and length of documents",
                "To identify important words in the dataset",
                "To measure the significance of individual words",
            ],
            "Pros": [
                "Improves data quality by removing irrelevant documents",
                "Filters out low-quality or incomplete documents",
                "Provides additional information for analysis",
                "Enables language-specific analysis and insights",
                "Helps understand the complexity and content of documents",
                "Identifies important terms and topics in the dataset",
                "Quantifies the importance of individual words",
            ],
            "Cons": [
                "May exclude documents in less common languages",
                "May remove documents with valuable information",
                "May introduce bias in the analysis",
                "May not accurately represent the language distribution",
                "May not capture the complexity of document structure",
                "May be sensitive to noise and outliers",
                "May not capture the semantic meaning of words",
            ],
        }
    )

    table_html = preprocessing_steps.to_html(index=False, border=0)
    table_div = Div(NotStr(table_html), style="margin: 40px;")
    data_preprocessing_div = Div(H3("Data Preprocessing"), text, table_div)



  

    return Div(
        Section(
            H2("Curated Sources: Overview"),
            overview_text,
            copyright_disclaimer,
            table_desc,
            H2("Curated Sources: Data Gathering and Filtering"),
            H3("Data Acquisition"),
            data_preparation_div,
            H3("Data Filtering"),
            data_preprocessing_div,
            plotly2fasthtml(get_chart_28168342()),
            H2("Local Deduplication"),
            table_div_data_pipe,
            id="inner-text",
        )
    )