Datasets:
Clarification on data deletion rules
I'm trying to understand how certain data is removed.
For example, for Fra_Latn_removed
- id: urn:uuid:2867ede1-5c90-424c-8b1a-1b5695dc0040
- url: http://archives.cg58.fr/?id=partager
- filter_reason: duplicated_6_n_grams
Where are the rules for such deletions defined?
Is there a specific part of the codebase (e.g., in Python) where these filtering criteria like duplicated_6_n_grams are implemented?
Any pointers to documentation or relevant source code would be appreciated!
Okay, I figured it out. The removal rules are defined in the filter steps of FineWeb 2's Python pipeline code (implemented via classes like GopherRepetitionFilter) and are configurable via YAML files in the repository (under the configs/ directory, e.g., configs/fra_Latn.yml). When a document (e.g., a French document) is flagged for containing too many duplicate 6-grams, the filter (defined in the French configuration) flags it with a reason (e.g., "duplicated_6_n_grams") and the document is stored separately ( (it is suffixed by _removed).