TextVQA vs The Stack v2

Provides a curated collection of 45K question-answer pairs paired with 28K images sourced from OpenImages, where questions require models to detect, recognize, and reason about text visible within image regions. The dataset architecture combines image-level annotations with character-level OCR ground truth, enabling training of end-to-end systems that jointly perform text detection, recognition, and semantic reasoning without pipeline decomposition.

Unique: Explicitly bridges OCR and VQA by requiring models to read text from images as a prerequisite for answering questions, rather than treating text as incidental; uses OpenImages as source material to ensure diverse real-world image contexts (documents, signs, product packaging, street scenes) rather than synthetic or controlled environments

vs alternatives: Differs from general VQA datasets (VQA v2, GQA) by making text reading a core requirement rather than optional, and from pure OCR datasets (ICDAR) by grounding text recognition in semantic question-answering tasks that measure practical utility

Provides standardized train/validation/test splits (45K questions across 28K images) with associated metrics infrastructure for measuring model accuracy on text-dependent visual reasoning. The evaluation framework enables comparison of end-to-end multimodal systems using metrics like accuracy, F1 score on OCR tokens, and answer-level correctness, supporting both pipeline and joint models through flexible annotation formats.

Unique: Evaluation framework explicitly measures the intersection of OCR and reasoning capabilities by requiring models to both detect/recognize text AND answer questions about it, rather than evaluating these as separate tasks; provides structured comparison across models with different OCR backends (learned vs. traditional)

vs alternatives: More rigorous than ad-hoc evaluation because it uses a fixed, large-scale benchmark with standardized splits, but less flexible than custom evaluation scripts that can measure task-specific metrics like OCR token-level F1 or reasoning accuracy in isolation

Defines a structured annotation format that pairs images with question-answer pairs and includes OCR ground truth (detected text, bounding boxes, character-level confidence scores). The schema supports multiple answer formats (free-form text, multiple choice, span selection) and enables training systems that learn to jointly optimize text detection, recognition, and semantic reasoning through end-to-end supervision.

Unique: Schema explicitly includes OCR ground truth (detected text, bounding boxes, confidence scores) as first-class annotations rather than auxiliary metadata, enabling models to learn text localization and recognition jointly with semantic reasoning; supports multiple answer formats (free-form, multiple choice) to accommodate different downstream task requirements

vs alternatives: More structured than raw image-question pairs because it includes OCR ground truth and bounding boxes, enabling pixel-level supervision; simpler than full scene graph annotations (Visual Genome) because it focuses narrowly on text understanding rather than comprehensive object and relationship labeling

Enables assessment of how models trained on TextVQA generalize to other vision-language tasks (e.g., general VQA, document understanding, scene text recognition) by providing standardized data splits and evaluation protocols. The framework supports transfer learning experiments where TextVQA serves as pretraining data or auxiliary task, measuring downstream performance on related benchmarks through unified metric computation.

Unique: Explicitly designed to measure transfer learning value of OCR-VQA pretraining by providing standardized evaluation protocols that isolate the contribution of text understanding to downstream tasks; enables systematic comparison of pretraining data mixtures (TextVQA-only, TextVQA + general VQA, etc.)

vs alternatives: More focused than general transfer learning benchmarks (VTAB, ImageNet) because it specifically measures OCR-VQA transfer value; more comprehensive than single-task evaluation because it tests generalization across multiple downstream tasks

Provides utilities for efficient sampling of image-question-answer triplets from the 45K questions across 28K images, supporting stratified sampling by question type, image domain, or answer length. The batching infrastructure handles variable-length sequences (questions, answers, OCR tokens) through padding/truncation and enables data augmentation (image crops, rotations) while preserving text visibility and semantic correctness.

Unique: Sampling and batching utilities are specifically designed for OCR-VQA by supporting stratification on text-related properties (OCR token count, text density in image) and augmentation strategies that preserve text readability; enables curriculum learning where models first learn simple text reading before complex reasoning

vs alternatives: More specialized than generic data loaders (PyTorch DataLoader) because it includes OCR-aware sampling and augmentation; more flexible than fixed batch construction because it supports dynamic stratification and curriculum learning strategies

A comprehensive dataset for training models on visual question answering, requiring the integration of OCR capabilities to interpret text within images, featuring 45K questions across 28K images.

Unique: This dataset specifically focuses on the challenge of integrating text recognition within visual contexts, setting it apart from standard visual datasets.

vs alternatives: Unlike other datasets, TextVQA uniquely combines visual and textual understanding, making it ideal for developing advanced OCR-integrated models.

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