CulturaX

Performs exact and fuzzy deduplication across 6.3 trillion tokens spanning 167 languages by combining mC4 and OSCAR source datasets with language-aware normalization and document-level hashing. Uses probabilistic data structures (likely Bloom filters or MinHash) to identify and remove duplicate content while preserving language-specific variations, reducing storage footprint and preventing model training on redundant examples that would skew learned distributions.

Applies multi-stage quality filtering using language-specific heuristics (character distributions, script validity, toxicity markers, repetition patterns) to remove low-quality documents before inclusion in the final dataset. Filters are tuned per-language family (Latin, CJK, Indic, etc.) to account for different character frequencies, punctuation norms, and valid repetition patterns, preventing models from learning from spam, gibberish, or machine-generated noise while preserving legitimate content in morphologically-rich languages.

Organizes 6.3 trillion tokens across 167 languages with explicit stratification, allowing users to sample or weight languages during training to balance representation and prevent high-resource languages (English, Chinese, Spanish) from dominating model behavior. Provides language-level metadata and sampling utilities so practitioners can construct training splits that reflect target deployment demographics rather than web-crawl frequency distributions, which are heavily skewed toward English and a few other high-resource languages.

Merges mC4 (English-heavy, 100+ languages, 750B tokens) and OSCAR (more balanced, 166 languages, 180B tokens) into a single unified corpus with consistent schema, metadata format, and access patterns through Hugging Face Datasets. Handles schema reconciliation, timestamp alignment, and source attribution so users can trace documents back to original crawls while treating the combined dataset as a single coherent resource, eliminating the need to manage two separate pipelines or worry about overlapping content.

Provides pre-computed statistics at token, document, and language levels (token counts per language, document length distributions, character set coverage, script family breakdown) accessible through Hugging Face Datasets metadata API. Enables practitioners to understand dataset composition without downloading the full corpus, supporting informed decisions about sampling strategies, language weighting, and expected model behavior across languages without requiring custom analysis scripts.

Integrates with Hugging Face Datasets library's streaming, caching, and distributed loading infrastructure, enabling efficient access patterns for training at scale. Supports streaming mode (load documents on-demand without downloading full corpus), local caching with automatic decompression, and distributed data loading across multiple GPUs/TPUs through Datasets' built-in sharding and sampling utilities, reducing memory footprint and enabling training on machines with limited disk space.

Enables streaming access to the 6.3 trillion token dataset without downloading the full corpus, using Hugging Face Datasets streaming mode to load documents on-the-fly during training. Supports batching, shuffling, and caching strategies optimized for distributed training pipelines to minimize memory footprint while maintaining training efficiency.

Automatically detects language for each document and normalizes text across diverse writing systems (Latin, Cyrillic, Arabic, CJK, Indic scripts, etc.) to ensure consistent preprocessing across all 167 languages. Uses language detection models (fastText or similar) with confidence thresholding and script-aware normalization (Unicode normalization, diacritic handling) to handle multilingual text robustly.

Computes multi-dimensional quality scores for each document based on content properties (text length, language detection confidence, character distribution, readability metrics) and metadata signals (domain reputation, crawl freshness, source reliability). Enables ranking and filtering documents by quality without binary accept/reject decisions, supporting nuanced quality-based sampling.

Analyzes document source domains (news sites, academic papers, social media, forums, etc.) and applies domain-specific filtering rules to balance representation across content types. Prevents domain-specific biases (e.g., over-representation of news or Wikipedia) that could skew model behavior toward particular writing styles or information sources.