Multilingual Corpus Deduplication At Scale

via “multilingual web corpus with consistent annotation across 5 languages”

30 trillion token web dataset with 40+ quality signals per document.

Unique: Provides 30 trillion tokens across 5 languages with identical quality signal annotations, enabling comparative studies of language-specific data characteristics and training multilingual models on a standardized base. Consistent annotation methodology across languages enables cross-language analysis.

vs others: Larger multilingual coverage (5 languages, 30 trillion tokens) than RedPajama-1T (English-only, 1 trillion tokens) and most competitors; consistent annotation enables comparative language research, but limited to European languages vs. competitors with broader language coverage.

via “multilingual-corpus-deduplication-at-scale”

6.3T token multilingual dataset across 167 languages.

Unique: Combines mC4 (English-heavy, 100+ languages) and OSCAR (more balanced, 166 languages) with unified deduplication pipeline, then applies language-aware normalization before hashing — most open datasets deduplicate within a single source, not across heterogeneous multilingual sources with different crawl dates and quality profiles

vs others: Larger and more language-inclusive than mC4 alone (6.3T vs 750B tokens) and more deduplicated than raw OSCAR, making it more suitable for training models that perform well across low-resource languages without overfitting to English-dominant patterns

via “multilingual parallel corpus discovery via searchable index”

Massive parallel corpus for machine translation.

Unique: Aggregates and indexes 1,214 distinct corpora from heterogeneous sources (subtitles, EU documents, web crawls, academic sources) into a unified searchable interface, rather than requiring users to visit individual corpus repositories. Maintains version tracking across releases (e.g., OpenSubtitles v2024 vs historical versions) and exposes corpus composition percentages relative to the full 102.9B sentence pair collection.

vs others: Broader corpus coverage (1,214 corpora, 1,005 languages) than single-source alternatives like OpenSubtitles alone, but lacks the quality filtering, alignment confidence scores, and API-based programmatic access that commercial MT platforms provide.

via “multilingual-text-corpus-extraction-from-web-crawl”

Multilingual web corpus covering 101 languages.

Unique: Processes Common Crawl at petabyte scale with language-aware segmentation across 101 languages, providing pre-filtered language-specific subsets rather than requiring downstream filtering. Uses probabilistic language ID to avoid expensive manual annotation while maintaining reasonable precision for high-resource languages.

vs others: Larger and more multilingual than OSCAR (85 languages) and more web-representative than Wikipedia-derived corpora, but with lower quality control than curated datasets like GLUE or SuperGLUE

via “multi-language code dataset curation with near-deduplication”

250GB curated code dataset for StarCoder training.

Unique: Applies probabilistic near-deduplication at scale across 86 languages with language-aware filtering, rather than simple string matching or language-agnostic hashing. Integrates GitHub issues and commits as additional code context, not just raw source files.

vs others: Larger and more diverse than CodeSearchNet (14 languages, 6M examples) and more aggressively deduplicated than raw The Stack, striking a balance between scale and training efficiency that Codex/GPT-4 datasets don't publicly expose.

via “sentence-level deduplication at scale”

Google's cleaned Common Crawl corpus used to train T5.

Unique: Applies sentence-level deduplication at scale across 750GB using deterministic techniques, removing redundant training examples while maintaining document structure; enables cleaner training data without requiring learned quality models

vs others: More thorough than document-level deduplication; simpler and more reproducible than semantic deduplication approaches; reduces training data size but may miss near-duplicates that learned methods would catch

via “multilingual information retrieval with language-agnostic ranking”

sentence-similarity model by undefined. 4,39,47,771 downloads.

Unique: Operates in a unified multilingual embedding space learned from 50+ languages simultaneously, enabling direct similarity comparison between queries and documents in different languages without intermediate translation or language-specific indices, unlike traditional IR systems that require separate indices per language

vs others: Eliminates need for language detection, translation pipelines, and separate indices per language, reducing infrastructure complexity and latency by 5-10x compared to translation-based retrieval while maintaining competitive ranking quality

via “paraphrase-mining-and-duplicate-detection”

Framework for sentence embeddings and semantic search.

Unique: Provides specialized paraphrase mining API optimized for large-scale corpus processing with vectorized similarity computation, avoiding naive O(n²) pairwise comparisons; differentiates from generic similarity tools by handling batch processing and threshold filtering internally for production-scale deduplication

vs others: More efficient than manual duplicate detection or regex-based approaches because it understands semantic similarity rather than string matching, and simpler than building custom mining pipelines with separate embedding and similarity computation steps

via “language-agnostic semantic clustering and deduplication”

sentence-similarity model by undefined. 70,32,108 downloads.

Unique: Leverages multilingual-e5-small's shared embedding space to cluster texts across 94 languages without language-specific preprocessing or translation. The model's contrastive training ensures semantically equivalent texts cluster together regardless of language, enabling language-agnostic deduplication and grouping.

vs others: More accurate than lexical deduplication (string matching, fuzzy matching) for semantic equivalence; faster than translation-based approaches; supports 94 languages in a single model vs. language-specific clustering pipelines.

via “cross-lingual semantic matching and retrieval”

sentence-similarity model by undefined. 24,53,432 downloads.

Unique: Trained on diverse multilingual parallel and comparable corpora with contrastive learning that explicitly aligns semantically equivalent sentences across language pairs, creating a unified embedding space where cross-lingual similarity is directly comparable without separate language-pair-specific models or pivot languages

vs others: Achieves 15-20% higher cross-lingual retrieval accuracy than mBERT-based approaches on MTEB multilingual benchmarks while supporting 100+ languages in a single model, compared to language-pair-specific models that require O(n²) separate models for n languages

via “document clustering and deduplication”

sentence-similarity model by undefined. 36,60,082 downloads.

Unique: Operates on multilingual embeddings in a unified space, enabling clustering that respects semantic similarity across languages rather than creating separate clusters for each language — a Spanish document about 'cars' clusters with an English document about 'automobiles' rather than with other Spanish documents

vs others: More accurate than TF-IDF or BM25-based clustering for semantic grouping, and requires no language-specific preprocessing unlike traditional NLP clustering pipelines

via “cross-lingual-semantic-matching”

feature-extraction model by undefined. 32,39,437 downloads.

Unique: Multilingual BERT backbone trained on 215M parallel sentence pairs creates a shared embedding space where semantic meaning is preserved across 50+ languages without language-specific adapters or separate models — enables true zero-shot cross-lingual retrieval by design rather than post-hoc translation

vs others: Outperforms language-agnostic approaches (e.g., translating everything to English) by preserving nuance and avoiding translation errors; more efficient than maintaining separate monolingual models per language while achieving comparable or better cross-lingual accuracy

via “multilingual web-scale text corpus ingestion and deduplication”

Dataset by allenai. 7,61,810 downloads.

Unique: C4 is built directly from Common Crawl snapshots with transparent, reproducible filtering and deduplication logic (published in the original paper), making it auditable and replicable — unlike proprietary datasets. It includes explicit language detection and URL-based quality filtering applied uniformly across 100+ languages, enabling fair multilingual representation.

vs others: C4 offers 10x larger scale and true multilingual coverage compared to English-only datasets like Wikipedia or BookCorpus, while maintaining open-source transparency and reproducibility that proprietary datasets (e.g., GPT-3's training data) cannot provide.

via “deduplication at document and near-duplicate levels”

Dataset by HuggingFaceFW. 6,43,166 downloads.

Unique: Applies both exact and near-duplicate deduplication at Common Crawl scale with explicit benchmark contamination prevention, ensuring evaluation integrity — most web corpora lack deduplication or benchmark-aware filtering

vs others: Prevents benchmark leakage that affects model evaluation fairness, whereas raw Common Crawl and many other corpora do not address this issue

via “semantic deduplication and near-duplicate detection”

Nomic's embedding model — semantic search and similarity — embedding model

Unique: Performs semantic deduplication without lexical matching, capturing paraphrases and translations that string-based methods miss. Local execution enables processing sensitive documents without external API calls.

vs others: More robust than hash-based or string-similarity deduplication for handling paraphrasing and translation; faster than manual review while maintaining semantic understanding unlike simple string matching.

via “deduplication and redundancy removal at scale”

Dataset by HuggingFaceFW. 4,14,812 downloads.

Unique: Applies document-level deduplication using scalable algorithms (likely MinHash or similar) across the full 3.5B token corpus during preprocessing, removing both exact and near-duplicate content before release. Deduplication is transparent to users but not configurable post-hoc.

vs others: More efficient for training than raw Common Crawl or unfiltered FineWeb because redundancy is pre-removed, reducing wasted compute on duplicate examples; more principled than ad-hoc deduplication in training scripts because it's applied consistently across the full corpus.

via “multi-source text corpus aggregation and deduplication”

Dataset by LLM360. 10,70,517 downloads.

Unique: Combines web, book, and academic sources with explicit deduplication as part of the LLM360 transparency initiative, making source composition auditable unlike black-box datasets; balances representation across domains rather than raw-crawling dominance

vs others: More transparent about deduplication and source composition than Common Crawl or C4 (which publish minimal filtering details); smaller but more curated than raw web crawls, trading scale for quality and auditability