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 from eval_result_figures import all_eval_res_figs 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 = { "title": "TxT360", "description": "A globally deduplicated dataset for LLM pretraining", "published": "October 7, 2024", "authors": [ { "author": "Liping Tang", "authorURL": "https://huggingface.co/Liping", "affiliation": "MBZUAI", "affiliationURL": "LLM360.ai", }, { "author": "Nikhil Ranjan", "authorURL": "https://huggingface.co/nikhilranjan", "affiliation": "MBZUAI", "affiliationURL": "LLM360.ai", }, { "author": "Omkar Pangarkar", "authorURL": "https://huggingface.co/omkarenator", "affiliation": "Petuum, Inc.", "affiliationURL": "", }, { "author": "Xuezhi Liang", "authorURL": "", "affiliation": "MBZUAI", "affiliationURL": "", }, { "author": "Zhen Wang", "authorURL": "", "affiliation": "MBZUAI", "affiliationURL": "", }, { "author": "Li An", "authorURL": "https://huggingface.co/an1118", "affiliation": "UCSD", "affiliationURL": "", }, { "author": "Bhaskar Rao", "authorURL": "", "affiliation": "MBZUAI", "affiliationURL": "", }, { "author": "Zhoujun Cheng", "authorURL": "https://huggingface.co/zhoujun", "affiliation": "UCSD", "affiliationURL": "", }, { "author": "Suqi Sun", "authorURL": "https://huggingface.co/mylibrar", "affiliation": "Petuum, Inc.", "affiliationURL": "", }, { "author": "Cun Mu", "authorURL": "https://huggingface.co/CarisMu", "affiliation": "MBZUAI", "affiliationURL": "", }, { "author": "Victor Miller", "authorURL": "https://huggingface.co/vamiller12", "affiliation": "Petuum, Inc.", "affiliationURL": "", }, { "author": "Yue Peng", "authorURL": "https://huggingface.co/Dreamever", "affiliation": "MBZUAI", "affiliationURL": "", }, { "author": "Eric P. Xing", "authorURL": "", "affiliation": "MBZUAI", "affiliationURL": "https://www.mbzuai.ac.ae/ & https://www.cs.cmu.edu/", }, { "author": "Zhengzhong Liu", "authorURL": "https://huggingface.co/hunterhector", "affiliation": "Petuum, Inc. / MBZUAI ", "affiliationURL": "", }, ], "katex": {"delimiters": [{"left": "$$", "right": "$$", "display": "false"}]}, } citation_long = """ @misc{txt360data2024, title = {TxT360: a globally deduplicated dataset for LLM pretraining}, author = {Liping Tang, Nikhil Ranjan, Omkar Pangarkar, Zhen Wang, An Li, Zhoujun Cheng, Suqi Sun, Cun Mu, Victor Miller, Yue Peng, Eric P. Xing, Zhengzhong Liu}, year = 2024 } """ 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(): from fasthtml.xtend import Script return Div( D_title( H1( "TxT360: A Top-Quality LLM Pre-training Dataset Requires the Perfect Blend", cls="l-body", style="text-align: center;", ), Div( Img(src="images/llm360_logo.png"), id="title-plot", cls="main-plot-container l-page", ), ), D_byline(), D_front_matter( Script( json.dumps(front_matter), id="distill-front-matter", type="text/json", ) ), D_article( D_contents( Nav( H3("Table of Contents"), Div( A( "TxT360", href="#section11", ) ), Div( Ul( Li( A( "About TxT360", href="#section11", ) ), Li( A( "Why TxT360", href="#section12", ) ), Li( A( "Generalizable Approach to Data Processing", href="#section13", ) ), ), ), Div( A( "Common Crawl Data", href="#section21", ) ), Div( Ul( Li( A( "Common Crawl Snapshot Processing", href="#section21", ) ), Li( A( "Common Crawl Data Processing Summary", href="#section22", ) ), Li( A( "Document Preparation", href="#section23", ) ), Li( A( "Line-Level Removal", href="#section24", ) ), Li( A( "Document-Level Filtering", href="#section25", ) ), ), ), Div( A( "Curated Sources", href="#section31", ) ), Div( Ul( Li( A( "Curated Sources in TxT360", href="#section31", ) ), Li( A( "Filtering Steps and Definitions", href="#section32", ) ), Li( A( "Filtering Discussion on All Curated Sources", href="#section33", ) ), ), ), Div( A( "Shared Processing Steps", href="#section41", ) ), Div( Ul( Li( A( "Overview", href="#section41", ) ), Li( A( "Why Global Deduplication", href="#section42", ) ), Li( A( "MinHash Generation", href="#section43", ) ), Li( A( "Matching Pairs Generation", href="#section44", ) ), Li( A( "Finding Duplicate Pairs", href="#section45", ) ), Li( A( "Finding Connected Components using MapReduce", href="#section46", ) ), Li( A( "Personally Identifiable Information Removal", href="#section47", ) ), Li( A( "Normalization Form C", href="#section48", ) ), ), ), Div( A( "TxT360 Studies", href="#section51", ), ), Div( Ul( Li( A( "Overview", href="#section51", ) ), Li( A( "A Simple Data Mix Creates a Good Learning Curve", href="#section52", ) ), Li( A( "Perplexity Analysis", href="#section53", ) ), Li( A( "Topic Analysis", href="#section55", ) ) ), ), role="navigation", cls="l-text figcaption", ), ), intro(), web.web_data(), curated.curated(), common.common_steps(), results.results(), ), D_appendix( D_bibliography(src="bibliography.bib"), H3("Citation"), P("For attribution in academic contexts, please cite this work as"), Pre(citation_long, cls="citation long"), ), Div(*read_bibs(), style="display: none;"), ) new_dataset_comparison1 = pd.DataFrame( { "Data Source": [ "CommonCrawl Snapshots", "Papers", "Wikipedia", "FreeLaw", "DM Math", "USPTO", "PG-19", "HackerNews", "Ubuntu IRC", "EuroParl", "StackExchange", "Code", ], "TxT360": [ "99", "5 Sources", "310+ Languages", "Included", "Included", "Included", "Included", "Included", "Included", "Included", "Included", "**", ], "FineWeb": [ "96", "-", "-", "-", "-", "-", "-", "-", "-", "-", "-", "-", ], "RefinedWeb": [ "90", "-", "-", "-", "-", "-", "-", "-", "-", "-", "-", "-", ], "PedPajamaV2": [ "84", "-", "-", "-", "-", "-", "-", "-", "-", "-", "-", "-", ], "C4": [ "1", "-", "-", "-", "-", "-", "-", "-", "-", "-", "-", "-", ], "Dolma": [ "24", "1 Source", "Included", "-", "-", "-", "Included", "-", "-", "-", "-", "Included", ], "RedPajamaV1": [ "5", "1 Source", "Included", "", " ", "", "Included", "-", "-", "-", "Included", "Included", ], "The Pile": [ "0.6% of 74", "4 Sources", "English Only", "Included", "Included", "Included", "Included", "Included", "Included", "Included", "Included", "Included", ], } ) styled_table = ( new_dataset_comparison1.style.applymap( lambda _: "background-color: #E1EEDB", # Green background for col 1 subset=pd.IndexSlice[:, "TxT360"], ) .applymap( lambda _: "background-color: white", # White background for all other columns subset=pd.IndexSlice[ :, new_dataset_comparison1.columns.difference(["TxT360"]) ], # Apply to all columns except "TxT360" ) .set_properties( **{ "text-align": "left", # Left the text in each cell "padding": "10px", # Add padding for better readability "word-wrap": "break-word", # Ensure text wraps within cells } ) .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) # new_table_div_1 = Div(NotStr(table_html), style="margin: 40px;") new_table_div_1 = Div( NotStr(table_html), style="display: flex; justify-content: center; align-items: center; width: 100%; max-width: 100%; height: auto; overflow-x: auto;" ) 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" # Green background for row 0 if i == 0 else "background-color: rgb(237, 242, 251)" # Blue background for other rows 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, ) .set_table_styles( [ {"selector": "table", "props": [("margin-left", "auto"), ("width", "100%")]}, # Make table responsive and centered ] ) .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;") table_div_2 = Div(NotStr(table_html2)) dataset_sources = pd.DataFrame( { "Data Source": [ "CommonCrawl", "Papers", "Wikipedia", "Freelaw", "DM Math", "USPTO", "PG-19", "HackerNews", "Ubuntu IRC", "Europarl", "StackExchange", ], "Raw Data Size": [ "9.2 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": [ "4.83T", "154.96B", "4.75B", "7.34B", "5.23B", "4.95B", "2.94B", "1.08B", "1.54B", "1.96B", "8.37B", ], "Information 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, ) .set_properties( **{ "text-align": "center", # Center the text in each cell "padding": "10px", # Add padding for better readability "word-wrap": "break-word", # Ensure text wraps within cells } ) .hide(axis="index") # Hide the row index ) table_html_data = styled_table._repr_html_() # Wrap the table in a Div, ensuring it is centered table_div_data = Div( NotStr(table_html_data), # style="margin-left: auto; width: 90%; max-width: 100%; text-align: center; align: center; overflow-x: auto;" style="display: flex; justify-content: center; align-items: center; width: 100%; max-width: 100%; height: auto; overflow-x: auto;" ) @app.get("/intro") def intro(): return Div( Section( H2("About TxT360"), P( B("TL;DR "), "We introduce ", A(B("TxT360 (Trillion eXtracted Text),"), href="https://huggingface.co/datasets/LLM360/TxT360"), " the first dataset to globally deduplicate 99 CommonCrawl snapshots and 14 high-quality data sources from diverse domains (e.g., FreeLaw, PG-19, etc.). The large-scale deduplication process and rich metadata stored enables precise control over data distribution. We demonstrate a simple but effective upsampling recipe that creates a 15+ trillion-token corpus, outperforming FineWeb 15T on several key metrics. With the information, TxT360 empowers pre-trainers to explore more advanced weighting techniques, a feature not commonly available in previous pre-training datasets. Our findings highlight the importance of both high-quality data sources and appropriate weighting for optimal blending in LLM training." ), P("In line with our 360° open source spirit, we document all detailed steps, reasons of our decisions, detailed statistics, our code (stay tuned!), analysis results and more, in additional to the dataset itself. We hope this can serve as a useful resource for future developers." ), plotly2fasthtml(all_eval_res_figs["MMLU"]), P( "Building on top of the prior studies on pre-training data", D_cite(bibtex_key="refinedweb"), D_cite(bibtex_key="fineweb"), D_cite(bibtex_key="c4"), D_cite(bibtex_key="muennighoff2023scaling"), D_cite(bibtex_key="dolma"), ", TxT360 carefully implements data processing steps including extraction, filtering, deduplication, personally identifiable information removal, and other steps. Unlike DCLM", D_cite(bibtex_key="dclm"), "and RedPajama V2,", D_cite(bibtex_key="redpajama-v2"), "we also hope to provide a dataset at this scale that is ready to go, without requiring futher filtering." ), id="section11", ), Section( H2("Why TxT360"), P( "In this year we have seen excellent datasets released by the community. Among those, most datasets focus on one source (e.g., crawled websites, code bases, papers). However, it is not trivial to combine these sources together due to the potential duplicaiton across them. TxT360 is the first dataset to combine most of sources commonly used in pretraining." ), new_table_div_1, # table_div_1, # table_div_2, P( "In LLM pretraining, it is common to combine all possible text sources due to the Scaling Law. Crawled web pages are included to provide a vast quantity of data which can cover long tail and diverse information, while curated datasets such as Wikipedia are also used, which often provide the 'deep-dive' domain information. 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." ), P( "** TxT360 does not include very specific domains such as code and math. This decision was made due to the perceived low duplication code with other sources, and the different logic requiring to build those datasets. We leave those work to future work and recommend users refer to existing projects such as Stack V2", D_cite(bibtex_key="lozhkov2024starcoder2stackv2"), ".", ), # P("Table 2: Basic TxT360 Statistics."), # table_div_data, id="section12", ), Section( H2("Our Approach"), P( "To produce TxT360, a comprehensive 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, but also can cause troubles with their own special formatting preferences. 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 ~5T tokens of high-quality data. The table below shows the source distribution of TxT360 tokens. ", B("Note that we do not recommend to use the raw distribution of the deduplicated dataset, a simple recipe is provided in the studies section."), ), table_div_data, P( "We provide details and context for the choices behind TxT360 in the respective Common Crawl Data Processing and Curated Source Processing section. A deep dive describing the deduplication process can be found in the Shared Processing Steps section." ), # 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="section13", ), id="inner-text", ) rt("/update/{target}")(data_viewer.update) serve()