ROOTS

DatasetFree

BigScience's curated multilingual dataset for BLOOM.

Open Source

/ 100

7 capabilities

Capabilities7 decomposed

multilingual pretraining corpus assembly with explicit language coverage

Medium confidence

ROOTS provides a curated collection of 46 natural languages and 13 programming languages organized into distinct data sources with documented provenance, enabling language-balanced pretraining without requiring custom data collection. The dataset uses a source-level organization pattern where each language's data is grouped by origin (web crawls, books, code repositories, etc.), allowing trainers to inspect and weight language contributions independently during model training.

Solves for

Train a multilingual language model without manually sourcing and validating data across dozens of languagesUnderstand the exact composition and origin of training data for a production language modelReproduce BLOOM's training setup or adapt it for similar multilingual modelsAudit language representation and identify potential biases in pretraining corpora

Best for

Research teams training multilingual foundation models

Organizations building non-English language models with transparency requirements

Teams reproducing or extending BLOOM-family models

Requires

Hugging Face Datasets library (datasets>=2.0)

Minimum 2TB free disk space for full dataset download

Python 3.8+

Limitations

Dataset is fixed and immutable — no mechanism for incremental updates or corrections post-release

Language coverage is unbalanced; some languages have significantly less data than others due to web availability

No built-in deduplication across sources — duplicate content may exist between web crawls and other sources

What makes it unique

Combines explicit data governance documentation (sourcing rationale, licensing, potential biases) with source-level granularity, allowing researchers to inspect and selectively use subsets rather than treating the corpus as a black box. This architectural choice prioritizes transparency over convenience.

vs alternatives

More transparent and auditable than Common Crawl-only datasets, with documented language selection rationale; more diverse than English-only corpora like The Pile, but smaller and more curated than raw web-scale datasets like C4

source-level data filtering and composition control

Medium confidence

ROOTS organizes data into discrete sources (e.g., 'Wikipedia', 'GitHub', 'Books', 'Web Crawl') that can be independently selected, weighted, or excluded during dataset loading. This enables trainers to construct custom training mixes without re-downloading or reprocessing the entire corpus, using Hugging Face Datasets' filtering and streaming APIs to apply source-based selection at load time.

Solves for

Create a custom training mix emphasizing code and technical documentation over general web textExclude low-quality sources or sources with licensing concerns from your training runExperiment with different source compositions to measure impact on model capabilitiesBalance language representation by selectively including/excluding sources per language

Best for

Researchers experimenting with data composition's effect on model quality

Teams with specific licensing or quality requirements that exclude certain sources

Organizations training domain-specific models (e.g., code-focused) from a multilingual base

Requires

Hugging Face Datasets library with streaming support

Understanding of ROOTS source taxonomy and composition

Python 3.8+

Limitations

Source-level filtering is coarse-grained — cannot filter within a source (e.g., exclude low-quality Wikipedia articles)

No built-in tools for measuring source quality or overlap; requires external analysis

Filtering decisions must be made at load time; cannot dynamically adjust during training without restarting

What makes it unique

Implements source-level composition as a first-class operation rather than post-hoc filtering, allowing researchers to reason about data provenance and make deliberate choices about which sources contribute to training. This is enforced through the dataset's hierarchical structure in Hugging Face Hub.

vs alternatives

More flexible than fixed-composition datasets like C4, but less granular than document-level filtering systems; enables reproducible data composition decisions without requiring custom preprocessing pipelines

language-stratified data access with per-language source documentation

Medium confidence

ROOTS structures data with language as a primary dimension, providing separate subsets for each of 46 languages plus 13 programming languages. Each language's data includes documentation of which sources contributed, their relative proportions, and known quality/bias characteristics, enabling language-specific analysis and informed decisions about language inclusion in multilingual training.

Solves for

Train a model with balanced representation across target languagesAnalyze whether a specific language has sufficient data quality for production useUnderstand which sources dominate a particular language's representationBuild language-specific models or fine-tune multilingual models with language-aware data weighting

Best for

Multilingual model developers who need language-aware data composition

Teams building models for low-resource languages and needing to understand data limitations

Researchers studying language representation in pretraining corpora

Requires

Language code mapping (ISO 639-1 or custom ROOTS language identifiers)

Hugging Face Datasets library

Understanding of target language's linguistic characteristics

Limitations

Language definitions are fixed (46 natural languages + 13 programming languages) — cannot add new languages or redefine language boundaries

Language-specific documentation is narrative, not machine-readable; requires manual interpretation for programmatic use

Data imbalance across languages is significant; some languages have orders of magnitude less data than others

What makes it unique

Treats language as a structural dimension of the dataset rather than a filtering criterion, with dedicated documentation per language covering sources, proportions, and known limitations. This enables language-aware training strategies that would be difficult with language-agnostic corpora.

vs alternatives

More language-aware than generic web-scale datasets; provides explicit documentation of language composition unlike mC4 or other derived multilingual corpora, enabling informed decisions about language inclusion

programming language corpus with source-specific code quality tiers

Medium confidence

ROOTS includes 13 programming languages sourced from GitHub, Stack Overflow, and other code repositories, with implicit quality stratification based on source (e.g., GitHub stars, Stack Overflow votes). The corpus preserves source metadata allowing trainers to filter by code quality signals without requiring custom code quality evaluation, enabling code-focused model training with quality control.

Solves for

Train a code generation model with balanced representation across programming languagesExclude low-quality code (e.g., Stack Overflow answers with negative votes) from trainingAnalyze which programming languages are well-represented in the pretraining corpusBuild code-specific models by selectively including code sources while excluding natural language

Best for

Teams training code generation or code understanding models

Researchers studying programming language representation in LLMs

Organizations building multilingual models with strong code capabilities

Requires

Hugging Face Datasets library

Programming language identifiers (e.g., 'python', 'javascript')

Optional: code parsing library for language-specific analysis

Limitations

Code quality filtering is implicit (based on source) rather than explicit; no fine-grained quality metrics provided

Programming language coverage is limited to 13 languages; less common languages are underrepresented

No deduplication of code snippets across sources; identical solutions may appear multiple times

What makes it unique

Includes programming languages as a first-class data dimension with source-based quality signals (GitHub stars, Stack Overflow votes) preserved in metadata, enabling quality-aware code training without requiring external code quality evaluation systems.

vs alternatives

More comprehensive than single-source code datasets (e.g., GitHub-only), with implicit quality signals; smaller but more curated than raw GitHub dumps, making it suitable for production model training

dataset streaming and partial loading for memory-constrained environments

Medium confidence

ROOTS integrates with Hugging Face Datasets' streaming API, allowing researchers to load and process data without downloading the entire corpus to disk. Streaming uses an iterator-based pattern where documents are fetched on-demand from the Hub, enabling training on machines with limited storage while maintaining full dataset access through network I/O.

Solves for

Train models on ROOTS without requiring terabytes of local storagePrototype model training quickly by streaming a subset of dataProcess data in a streaming fashion to reduce memory footprint during preprocessingDistribute data loading across multiple machines without duplicating the full dataset

Best for

Researchers with limited local storage or budget constraints

Teams prototyping models before committing to full-scale training

Organizations using cloud training with per-GB storage costs

Requires

Hugging Face Datasets library with streaming support (>=2.0)

Stable network connection with sufficient bandwidth (10+ Mbps recommended)

Python 3.8+

Limitations

Streaming introduces network latency; throughput is limited by network bandwidth and Hub availability

Cannot efficiently shuffle or reorder data when streaming; random access is not supported

Streaming performance degrades with large batch sizes or high-concurrency access

What makes it unique

Leverages Hugging Face Datasets' streaming infrastructure to enable on-demand data access without local storage, using an iterator-based pattern that integrates seamlessly with PyTorch DataLoaders and distributed training frameworks.

vs alternatives

More storage-efficient than downloading full datasets; comparable to other Hub-hosted datasets but with better documentation and integration for multilingual training workflows

data governance and licensing metadata with per-source attribution

Medium confidence

ROOTS includes explicit licensing information and sourcing documentation for each data source, stored as structured metadata alongside the corpus. This enables automated license compliance checking and attribution generation, allowing trainers to verify that their training mix respects licensing constraints and to generate proper attribution statements for model cards.

Solves for

Verify that your training data mix complies with open-source and commercial licensing requirementsGenerate attribution statements for model cards documenting data sourcesIdentify sources with restrictive licenses that may limit model deploymentAudit data governance decisions made during model training

Best for

Organizations training models for commercial deployment with legal/compliance requirements

Research teams publishing models and needing to document data provenance

Teams building models with explicit data governance policies

Requires

Access to ROOTS dataset documentation and metadata

Understanding of open-source licenses (CC-BY, CC-BY-SA, MIT, etc.)

Optional: legal review for commercial deployment

Limitations

Licensing metadata is provided by BigScience but may not be exhaustive or legally authoritative; legal review is still required

No automated license compatibility checking; determining whether licenses conflict requires manual analysis

Attribution metadata is narrative, not machine-readable; generating compliant attribution requires manual effort

What makes it unique

Provides explicit per-source licensing and governance documentation as a first-class dataset feature, rather than burying it in README files. This enables programmatic license compliance checking and reproducible attribution generation.

vs alternatives

More transparent than datasets with minimal licensing information; comparable to other BigScience datasets but more comprehensive than typical web-scale corpora which lack detailed licensing metadata

community-curated data quality annotations and bias documentation

Medium confidence

ROOTS includes community-contributed annotations documenting known biases, quality issues, and limitations in specific sources, stored as structured metadata. These annotations are curated by BigScience and the research community, providing qualitative assessments of data quality and potential harms that complement quantitative metrics, enabling informed decisions about source inclusion.

Solves for

Understand known biases and limitations in specific data sources before including them in trainingMake informed decisions about excluding sources with documented quality or ethical concernsDocument known limitations of your model's training data in model cardsContribute quality annotations for sources you've analyzed

Best for

Teams building models with explicit bias and fairness considerations

Researchers studying bias in pretraining corpora

Organizations with ethical AI governance requirements

Requires

Access to ROOTS documentation and metadata

Understanding of bias types and fairness concepts

Critical reading skills for interpreting qualitative annotations

Limitations

Bias annotations are qualitative and subjective; no standardized bias metrics provided

Coverage is incomplete; not all sources have detailed bias documentation

Annotations reflect BigScience's perspective and may not capture all relevant concerns

What makes it unique

Incorporates community-curated bias and quality annotations as dataset metadata, treating data governance as an ongoing collaborative process rather than a one-time curation effort. This enables researchers to make informed decisions about data inclusion based on documented concerns.

vs alternatives

More transparent about known biases than datasets with minimal documentation; enables bias-aware training unlike datasets that treat data as neutral. Comparable to other BigScience datasets but with more extensive community input.

Capabilities are decomposed by AI analysis. Each maps to specific user intents and improves with match feedback.

Related Artifactssharing capabilities

Artifacts that share capabilities with ROOTS, ranked by overlap. Discovered automatically through the match graph.

Model44

MAP-Neo

Fully open bilingual model with transparent training.

bilingual data collection and preprocessingend-to-end transparent llm training pipeline

2 shared capabilities

Dataset46

RedPajama v2

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

language-specific corpus extraction and analysismultilingual web-scale pretraining corpus provision

2 shared capabilities

Dataset48

StarCoder Data

783 GB curated code dataset from 86 languages with PII redaction.

language-stratified corpus organization with balanced samplingmulti-language code corpus assembly with permissive licensing filtering

2 shared capabilities

Dataset26

c4

Dataset by allenai. 6,98,456 downloads.

language detection and multilingual corpus stratificationmultilingual web-scale text corpus ingestion and deduplication

2 shared capabilities

Dataset45

OPUS

Massive parallel corpus for machine translation.

multilingual parallel sentence alignment and retrievaldomain-stratified corpus filtering and sampling

2 shared capabilities

Dataset46

Dolma

Allen AI's 3T token dataset for fully reproducible LLM training.

multi-source pretraining corpus assembly with documented curation

1 shared capability

Best For

✓Research teams training multilingual foundation models
✓Organizations building non-English language models with transparency requirements
✓Teams reproducing or extending BLOOM-family models
✓Researchers experimenting with data composition's effect on model quality
✓Teams with specific licensing or quality requirements that exclude certain sources
✓Organizations training domain-specific models (e.g., code-focused) from a multilingual base
✓Multilingual model developers who need language-aware data composition
✓Teams building models for low-resource languages and needing to understand data limitations

Known Limitations

⚠Dataset is fixed and immutable — no mechanism for incremental updates or corrections post-release
⚠Language coverage is unbalanced; some languages have significantly less data than others due to web availability
⚠No built-in deduplication across sources — duplicate content may exist between web crawls and other sources
⚠Requires substantial storage (terabytes) and compute for full preprocessing and tokenization
⚠Source-level filtering is coarse-grained — cannot filter within a source (e.g., exclude low-quality Wikipedia articles)
⚠No built-in tools for measuring source quality or overlap; requires external analysis

Requirements

Hugging Face Datasets library (datasets>=2.0)Minimum 2TB free disk space for full dataset downloadPython 3.8+Network bandwidth for multi-hour downloads depending on connection speedHugging Face Datasets library with streaming supportUnderstanding of ROOTS source taxonomy and compositionLanguage code mapping (ISO 639-1 or custom ROOTS language identifiers)Hugging Face Datasets library

Input / Output

Accepts: dataset identifier (bigscience-data/roots), language selection parameters, source filtering criteria, source names (string list), language codes (ISO 639-1 or custom), filtering predicates (optional), language code or name, optional source filter, optional date range (if available), programming language name or identifier, optional source filter (GitHub, Stack Overflow, etc.), optional quality threshold (implicit via source), dataset identifier, streaming configuration (batch size, buffer size), optional split specification, source name or identifier, optional license filter, optional bias category filter

Produces: raw text documents, structured metadata (source, language, date), tokenized sequences (when processed), filtered dataset subset, metadata about selected sources, composition statistics, language-specific text corpus, source composition metadata, language documentation (text), code snippets with metadata, source attribution, language-specific statistics, iterator over documents, batched tensors (when used with PyTorch DataLoader), metadata per document, license identifier (SPDX format or text), source attribution text, licensing constraints and restrictions, bias annotations (text), quality assessments, recommendations for source inclusion/exclusion

UnfragileRank

Adoption70%(35% weight)

Quality23%(25% weight)

Ecosystem40%(20% weight)

Match Graph10%(15% weight)

Freshness100%(5% weight)

UnfragileRank is computed from adoption signals, documentation quality, ecosystem connectivity, match graph feedback, and freshness. No artifact can pay for a higher rank.

Type: Dataset

7 capabilities

Visit ROOTS→

About

BigScience's curated multilingual dataset used to train BLOOM, covering 46 natural languages and 13 programming languages with explicit data governance, sourcing documentation, and community-driven curation.

Alternatives to ROOTS

cua53Agent

Open-source infrastructure for Computer-Use Agents. Sandboxes, SDKs, and benchmarks to train and evaluate AI agents that can control full desktops (macOS, Linux, Windows).

Compare →

Hugging Face43Platform

The GitHub for AI — 500K+ models, datasets, Spaces, Inference API, hub for open-source AI.

Compare →

Stable-Diffusion55Repository

FLUX, Stable Diffusion, SDXL, SD3, LoRA, Fine Tuning, DreamBooth, Training, Automatic1111, Forge WebUI, SwarmUI, DeepFake, TTS, Animation, Text To Video, Tutorials, Guides, Lectures, Courses, ComfyUI, Google Colab, RunPod, Kaggle, NoteBooks, ControlNet, TTS, Voice Cloning, AI, AI News, ML, ML News,

Compare →

YOLOv846Model

Real-time object detection, segmentation, and pose.

Compare →

Are you the builder of ROOTS?

Claim this artifact to get a verified badge, access match analytics, see which intents users search for, and manage your listing.

Claim this artifact →Verification via email

Get the weekly brief

New tools, rising stars, and what's actually worth your time. No spam.

Data Sources

seed developer essentials

Looking for something else?

Search →

Capabilities7 decomposed

multilingual pretraining corpus assembly with explicit language coverage

Medium confidence

Solves for

Best for

Research teams training multilingual foundation models

Organizations building non-English language models with transparency requirements

Teams reproducing or extending BLOOM-family models

Requires

Hugging Face Datasets library (datasets>=2.0)

Minimum 2TB free disk space for full dataset download

Python 3.8+

Limitations

Dataset is fixed and immutable — no mechanism for incremental updates or corrections post-release

Language coverage is unbalanced; some languages have significantly less data than others due to web availability

No built-in deduplication across sources — duplicate content may exist between web crawls and other sources

What makes it unique

vs alternatives

source-level data filtering and composition control

Medium confidence

Solves for

Best for

Researchers experimenting with data composition's effect on model quality

Teams with specific licensing or quality requirements that exclude certain sources

Organizations training domain-specific models (e.g., code-focused) from a multilingual base

Requires

Hugging Face Datasets library with streaming support

Understanding of ROOTS source taxonomy and composition

Python 3.8+

Limitations

Source-level filtering is coarse-grained — cannot filter within a source (e.g., exclude low-quality Wikipedia articles)

No built-in tools for measuring source quality or overlap; requires external analysis

Filtering decisions must be made at load time; cannot dynamically adjust during training without restarting

What makes it unique

vs alternatives

language-stratified data access with per-language source documentation

Medium confidence

Solves for

Best for

Multilingual model developers who need language-aware data composition

Teams building models for low-resource languages and needing to understand data limitations

Researchers studying language representation in pretraining corpora

Requires

Language code mapping (ISO 639-1 or custom ROOTS language identifiers)

Hugging Face Datasets library

Understanding of target language's linguistic characteristics

Limitations

Language definitions are fixed (46 natural languages + 13 programming languages) — cannot add new languages or redefine language boundaries

Language-specific documentation is narrative, not machine-readable; requires manual interpretation for programmatic use

Data imbalance across languages is significant; some languages have orders of magnitude less data than others

What makes it unique

vs alternatives

programming language corpus with source-specific code quality tiers

Medium confidence

Solves for

Best for

Teams training code generation or code understanding models

Researchers studying programming language representation in LLMs

Organizations building multilingual models with strong code capabilities

Requires

Hugging Face Datasets library

Programming language identifiers (e.g., 'python', 'javascript')

Optional: code parsing library for language-specific analysis

Limitations

Code quality filtering is implicit (based on source) rather than explicit; no fine-grained quality metrics provided

Programming language coverage is limited to 13 languages; less common languages are underrepresented

No deduplication of code snippets across sources; identical solutions may appear multiple times

What makes it unique

vs alternatives

dataset streaming and partial loading for memory-constrained environments

Medium confidence

Solves for

Best for

Researchers with limited local storage or budget constraints

Teams prototyping models before committing to full-scale training

Organizations using cloud training with per-GB storage costs

Requires

Hugging Face Datasets library with streaming support (>=2.0)

Stable network connection with sufficient bandwidth (10+ Mbps recommended)

Python 3.8+

Limitations

Streaming introduces network latency; throughput is limited by network bandwidth and Hub availability

Cannot efficiently shuffle or reorder data when streaming; random access is not supported

Streaming performance degrades with large batch sizes or high-concurrency access

What makes it unique

vs alternatives

More storage-efficient than downloading full datasets; comparable to other Hub-hosted datasets but with better documentation and integration for multilingual training workflows

data governance and licensing metadata with per-source attribution

Medium confidence

Solves for

Best for

Organizations training models for commercial deployment with legal/compliance requirements

Research teams publishing models and needing to document data provenance

Teams building models with explicit data governance policies

Requires

Access to ROOTS dataset documentation and metadata

Understanding of open-source licenses (CC-BY, CC-BY-SA, MIT, etc.)

Optional: legal review for commercial deployment

Limitations

Licensing metadata is provided by BigScience but may not be exhaustive or legally authoritative; legal review is still required

No automated license compatibility checking; determining whether licenses conflict requires manual analysis

Attribution metadata is narrative, not machine-readable; generating compliant attribution requires manual effort

What makes it unique

vs alternatives

More transparent than datasets with minimal licensing information; comparable to other BigScience datasets but more comprehensive than typical web-scale corpora which lack detailed licensing metadata

community-curated data quality annotations and bias documentation

Medium confidence

Solves for

Best for

Teams building models with explicit bias and fairness considerations

Researchers studying bias in pretraining corpora

Organizations with ethical AI governance requirements

Requires

Access to ROOTS documentation and metadata

Understanding of bias types and fairness concepts

Critical reading skills for interpreting qualitative annotations

Limitations

Bias annotations are qualitative and subjective; no standardized bias metrics provided

Coverage is incomplete; not all sources have detailed bias documentation

Annotations reflect BigScience's perspective and may not capture all relevant concerns

What makes it unique

vs alternatives

Capabilities are decomposed by AI analysis. Each maps to specific user intents and improves with match feedback.

Alternatives to ROOTS

cua53Agent

Open-source infrastructure for Computer-Use Agents. Sandboxes, SDKs, and benchmarks to train and evaluate AI agents that can control full desktops (macOS, Linux, Windows).

Compare →

Hugging Face43Platform

The GitHub for AI — 500K+ models, datasets, Spaces, Inference API, hub for open-source AI.

Compare →

Stable-Diffusion55Repository

Compare →

YOLOv846Model

Real-time object detection, segmentation, and pose.

Compare →

ROOTS

Capabilities7 decomposed

multilingual pretraining corpus assembly with explicit language coverage

source-level data filtering and composition control

language-stratified data access with per-language source documentation

programming language corpus with source-specific code quality tiers

dataset streaming and partial loading for memory-constrained environments

data governance and licensing metadata with per-source attribution

community-curated data quality annotations and bias documentation

Related Artifactssharing capabilities

MAP-Neo

RedPajama v2

StarCoder Data

c4

OPUS

Dolma

Best For

Known Limitations

Requirements

Input / Output

UnfragileRank

About

Categories

Alternatives to ROOTS

Are you the builder of ROOTS?

Get the weekly brief

Data Sources

ROOTS

Capabilities7 decomposed

multilingual pretraining corpus assembly with explicit language coverage

source-level data filtering and composition control

language-stratified data access with per-language source documentation

programming language corpus with source-specific code quality tiers

dataset streaming and partial loading for memory-constrained environments

data governance and licensing metadata with per-source attribution

community-curated data quality annotations and bias documentation

Related Artifactssharing capabilities

MAP-Neo

RedPajama v2

StarCoder Data

c4

OPUS

Dolma

Best For

Known Limitations

Requirements

Input / Output

UnfragileRank

About

Categories

Alternatives to ROOTS

Are you the builder of ROOTS?

Get the weekly brief

Data Sources