Visual Genome vs The Stack v2

Extracts and structures semantic relationships between objects in images using scene graph representations where nodes are objects and edges encode spatial/semantic relationships (e.g., 'person sitting on bench', 'cup on table'). The dataset provides pre-annotated scene graphs for 108K images, enabling models to learn structured reasoning about object interactions rather than treating images as flat feature vectors. Each relationship is labeled with predicate types (spatial: 'on', 'under'; semantic: 'wearing', 'holding') and grounded to pixel coordinates.

Unique: Provides densely annotated scene graphs at scale (2.3M relationships across 108K images) with explicit predicate types and pixel-level grounding, enabling structured learning of visual relationships rather than implicit feature-based representations. Uses hierarchical annotation combining object-level, attribute-level, and relationship-level labels in a unified graph structure.

vs alternatives: Richer than COCO (object detection only) and more structured than ImageNet (no relationship annotations); enables training models that reason about object interactions, not just recognition

Provides 5.4 million natural language descriptions grounded to specific image regions (bounding boxes), enabling training of vision-language models that map text to visual regions. Each region description is manually written by annotators and linked to pixel coordinates, creating a dense supervision signal for learning region-text alignment. Descriptions range from simple object names to complex compositional descriptions capturing attributes, actions, and relationships.

Unique: Provides 5.4M region descriptions with pixel-level grounding across 108K images, creating dense supervision for learning fine-grained region-text alignment. Uses multi-annotator consensus for quality control and covers diverse object categories, attributes, and compositional descriptions.

vs alternatives: Denser and more diverse than Flickr30K (158K descriptions) and provides explicit region coordinates unlike raw image-caption pairs; enables training region-grounding models at scale

Contains 1.7 million visual question-answer pairs grounded in scene context, where questions reference objects, relationships, and attributes visible in images. Questions are paired with images and scene graphs, enabling models to learn to answer questions by reasoning over visual structure rather than pattern-matching. Answer types range from simple object names to complex compositional answers requiring multi-step reasoning over relationships.

Unique: Integrates 1.7M QA pairs with scene graph annotations, enabling models to learn reasoning over structured visual knowledge rather than image-level features alone. Questions are grounded in specific objects and relationships, creating a tighter coupling between language and visual structure.

vs alternatives: Larger and more structured than VQA v2 (1.1M questions) and includes scene graph grounding unlike standard VQA datasets; enables training models that reason over visual relationships

Provides 3.8 million annotated object instances with bounding boxes, class labels, and 2.8 million attribute annotations (e.g., color, material, size, state). Each object is labeled with multiple attributes describing its visual properties, enabling training of models that predict not just object categories but fine-grained visual properties. Attributes are structured as key-value pairs (e.g., 'color: red', 'material: wood') and grounded to specific object instances.

Unique: Combines 3.8M object instances with 2.8M attribute annotations in a unified dataset, enabling training of attribute-aware detection models. Attributes are structured as key-value pairs and grounded to specific instances, creating dense supervision for learning visual properties beyond category labels.

vs alternatives: Richer attribute annotations than COCO (which has minimal attributes) and larger scale than fine-grained datasets like CUB-200 (11K images); enables training attribute-aware detection at scale

Integrates images, scene graphs, region descriptions, object attributes, and QA pairs into a unified multimodal dataset, enabling end-to-end training of vision-language models that learn from multiple supervision signals simultaneously. The dataset structure allows models to leverage complementary annotations (e.g., region descriptions for grounding, scene graphs for reasoning, attributes for fine-grained understanding) in a single training pipeline. Supports multi-task learning where models jointly optimize for detection, grounding, VQA, and relationship prediction.

Unique: Provides unified integration of 5 complementary annotation types (scene graphs, region descriptions, object instances, attributes, QA pairs) across 108K images, enabling multi-task learning from diverse supervision signals. Dataset structure supports joint optimization for detection, grounding, reasoning, and attribute prediction in a single training pipeline.

vs alternatives: More comprehensive than single-task datasets (COCO, Flickr30K) and enables multi-task learning unlike datasets with isolated annotation types; supports training unified models that leverage complementary supervision signals

Enables indexing and retrieval of images based on scene graph structure and relationships, allowing queries like 'find images with a person sitting on a bench' or 'images where a dog is next to a car'. Scene graphs are indexed as structured knowledge representations, supporting semantic search over visual relationships rather than keyword matching. Retrieval can be performed by querying for specific objects, relationships, or relationship patterns.

Unique: Provides 2.3M annotated relationships indexed as scene graphs, enabling structured retrieval by visual relationships and spatial configurations. Supports querying by relationship patterns (e.g., 'X on Y') rather than keyword matching, enabling semantic search over visual structure.

vs alternatives: Enables relationship-based retrieval unlike keyword-based image search; supports complex spatial/semantic queries that text-based systems cannot express

Provides statistical analysis and distribution information about visual relationships, objects, and attributes across the dataset, enabling researchers to understand frequency patterns, co-occurrence statistics, and relationship distributions. Includes statistics on predicate frequencies, object co-occurrence patterns, attribute distributions, and relationship types. Enables analysis of visual knowledge biases and patterns in the dataset.

Unique: Provides comprehensive statistical analysis of 2.3M relationships, 3.8M objects, and 2.8M attributes across 108K images, enabling researchers to understand visual knowledge distributions and dataset biases. Includes frequency statistics, co-occurrence patterns, and relationship type distributions.

vs alternatives: Enables large-scale statistical analysis of visual relationships unlike smaller datasets; provides insights into relationship distributions and biases for improving model training

Enables training of compositional visual understanding models by providing structured annotations that decompose images into objects, attributes, and relationships. Models can learn to compose understanding from parts (objects + attributes + relationships) rather than treating images as monolithic wholes. Supports learning of compositional generalization where models understand novel combinations of known objects and relationships.

Unique: Provides explicit decomposition of images into objects, attributes, and relationships, enabling training of compositional models that understand visual scenes through structured components. Scene graphs naturally support compositional learning by representing images as compositions of objects and relationships.

vs alternatives: Enables compositional learning unlike flat image-label datasets; supports training models that generalize to novel combinations of known components

Aggregates 67 TB of source code from the Software Heritage archive, filtering for permissively licensed repositories (MIT, Apache 2.0, BSD, etc.) across 600+ programming languages. Uses automated license detection and validation to ensure legal compliance for model training. Implements a rigorous deduplication pipeline at file and repository levels to eliminate redundant training data and reduce dataset bloat.

Unique: Largest open-source code dataset at 67 TB with automated opt-out governance allowing repository owners to request removal, combined with rigorous deduplication and PII removal pipeline — no other public dataset offers this scale with legal compliance and community control mechanisms

vs alternatives: Larger and more legally compliant than GitHub's CodeSearchNet (14M files) or Google's BigQuery public datasets, with explicit opt-out governance vs. implicit inclusion, and covers 600+ languages vs. Codex training data's undisclosed language distribution

Implements a community-driven opt-out system where repository owners can request removal of their code from the dataset without legal takedown notices. Maintains a registry of excluded repositories and re-applies exclusions during dataset updates. Provides transparent governance documentation and a clear submission process for removal requests, balancing open access with creator rights.

Unique: First large-scale code dataset to implement opt-out governance at dataset level rather than relying solely on license compliance, with transparent registry and community submission process — shifts power from dataset creators to code contributors

vs alternatives: More respectful of creator autonomy than GitHub Copilot's training approach (no opt-out) or academic datasets (one-time snapshot), and more scalable than individual DMCA takedowns

Automated pipeline that scans source code for personally identifiable information (email addresses, API keys, SSH keys, credit card patterns, phone numbers) and removes or redacts them before dataset release. Uses regex patterns, entropy-based detection for secrets, and heuristic rules to identify sensitive data. Operates at file level with configurable sensitivity thresholds to balance data utility against privacy risk.

Unique: Combines regex pattern matching, entropy-based secret detection, and heuristic rules in a unified pipeline with configurable sensitivity — more comprehensive than simple regex-only approaches, but trades off false positive rate against security coverage

vs alternatives: More thorough than GitHub's secret scanning (which only flags known patterns) because it includes entropy-based detection for unknown secret formats, but less accurate than specialized tools like TruffleHog due to language-agnostic approach

Indexes 67 TB of source code across 600+ programming languages with language-aware metadata (syntax, file extension, language family). Enables retrieval by language, license, repository, or code patterns. Uses Software Heritage's existing indexing infrastructure as foundation, augmented with language detection and classification. Supports both bulk download and filtered queries for specific language subsets.

Unique: Leverages Software Heritage's existing language detection and indexing infrastructure, then augments with BigCode-specific language classification and filtering — avoids reinventing language detection while providing dataset-specific query capabilities

vs alternatives: More comprehensive language coverage (600+ languages) than GitHub's Linguist (500+ languages) and more accessible than Software Heritage's raw API because it's pre-filtered for permissive licenses and deduplicated

Removes duplicate code files and repositories using content hashing (SHA-256 or similar) and fuzzy matching for near-duplicates. Operates in two stages: exact deduplication via hash matching, then fuzzy matching (e.g., Jaccard similarity or MinHash) to catch semantically identical code with minor formatting differences. Preserves one canonical copy of each unique code pattern while removing redundant training examples.

Unique: Two-stage deduplication combining exact hash matching with fuzzy similarity matching (likely MinHash or Jaccard) to catch both identical and near-identical code — more thorough than single-stage approaches but computationally expensive

vs alternatives: More aggressive deduplication than CodeSearchNet (which uses simple hash matching) because it catches near-duplicates, but less semantic than clone detection tools (which understand code structure) because it's content-based

Integrates with Software Heritage's comprehensive archive of 200+ million repositories and their full version control history. Extracts source code snapshots from Software Heritage's Git/Mercurial/SVN repositories, preserving repository metadata (commit history, author info, timestamps). Provides access to code at specific points in time, enabling historical analysis or training on code evolution patterns.

Unique: Leverages Software Heritage's universal code archive (200M+ repositories) as data source, providing access to code that would be impossible to collect via GitHub API alone — enables training on archived/deleted repositories and non-GitHub platforms (GitLab, Gitea, etc.)

vs alternatives: More comprehensive than GitHub-only datasets because it includes code from GitLab, Gitea, SourceForge, and other platforms archived by Software Heritage; more legally defensible than web scraping because it uses an established, community-maintained archive

Tracks and validates SPDX license identifiers for each repository, ensuring only permissively licensed code (MIT, Apache 2.0, BSD, etc.) is included. Maintains license metadata alongside code files, enabling downstream users to verify legal compliance. Implements license hierarchy and compatibility checking to handle dual-licensed or complex licensing scenarios.

Unique: Combines automated SPDX detection with manual review and maintains license metadata alongside code, enabling downstream users to verify compliance — more transparent than datasets that simply claim 'permissive licenses' without proof

vs alternatives: More legally rigorous than GitHub's CodeSearchNet (which doesn't validate licenses) and more transparent than Codex training data (which doesn't disclose license filtering at all)

Maintains versioned snapshots of the dataset (e.g., v2.0, v2.1) with documented changes between versions (new repositories added, deduplication improvements, PII removal updates). Provides checksums and manifests for reproducibility, enabling researchers to cite specific dataset versions and reproduce results. Tracks dataset lineage and transformation history.

Unique: Maintains semantic versioning and detailed changelogs for dataset releases, enabling researchers to cite specific versions and understand dataset evolution — more rigorous than one-off dataset releases without versioning

vs alternatives: More reproducible than academic datasets that are released once without versioning, and more transparent than commercial datasets (Codex) that don't disclose version history or changes