Hugging Face's logo

Datasets:
HuggingFaceFW
/
fineweb

Tasks:
Text Generation
Modalities:
Tabular
Text
Formats:
parquet
Languages:
English
Size:
10B - 100B
ArXiv:
DOI:
Libraries:
Datasets
Dask
License:
Dataset card Data Studio Files Community
59

Questions on Deduplication Strategy, Temporal Metadata, and Document Types

#62
by AshleyLL - opened
Discussion
AshleyLL

I hope you're all doing well. I'm reaching out to better understand some aspects of the FineWeb dataset, especially regarding deduplication strategies, temporal metadata, and content categorization. Any insights from the community would be greatly appreciated.

  1. Temporal Deduplication
    Does FineWeb perform deduplication across different time points? For example:
    If a webpage with the same URL (or near-identical content) appears in multiple years (e.g., 2013, 2014, 2015), is it deduplicated?
    If so:
    Which version is retained by default — earliest, latest, or highest-quality snapshot?
    Are there specific tools or metrics used to evaluate content equivalence over time?

  2. Proposed Deduplication Strategy
    We’re considering implementing a deduplication strategy based on the following rules:
    (1) Default Rule: Retain only the first occurrence of a webpage.
    (2) Exception Rule: Keep a subsequent crawl if:

  • The content has undergone significant modification (e.g., expanded depth/breadth), or
  • The new version is of higher quality.

Could the community share thoughts on this approach?
Does this align with best practices?
Are there known pitfalls or alternative strategies we should consider?
Has something like this already been implemented or is it under development in FineWeb?
If not, are there recommended tools or toolchains for implementing such a strategy at scale?
How can one define "significant content difference" or "higher quality"? E.g.:

  • Semantic similarity thresholds?
  • Content length or structural changes?
  • Quality heuristics (e.g., readability, domain authority)?

  1. Document Categories
    Lastly, could anyone confirm whether FineWeb includes documents from the following categories, and how they are identified or tagged?
    (1) Academic papers, including:
    Computer Science
    Mathematics, Physics, Medical Sciences
    Economics and Finance
    Other Humanities and Social Sciences
    (2) News articles, particularly:
    Financial and Economic news
    Political and Societal news
    Other news categories
    (3) Textbooks
    (4) Mathematical logic and reasoning, such as:
    Logical deduction problems
    Mathematical proofs or derivations
    Symbolic reasoning tasks
    Formal logic expressions or inference
    Any information on how these types are represented or labeled in the dataset would be extremely helpful.

  2. Temporal Metadata
    Could someone also confirm whether the dataset includes timestamps or modification history for each entry, such as:
    Original creation or generation time
    Update or revision timestamps
    Versioning information or change logs

Thank you all very much for your time and contributions to this project. Looking forward to hearing your thoughts!

Sign up or log in to comment