mmlu vs The Stack v2

Loads a structured dataset of 439,045 multiple-choice questions across 57 academic subjects (STEM, humanities, social sciences) created by expert annotators. The dataset is distributed via HuggingFace's datasets library in Parquet format with standardized schema (question, choices A-D, correct answer, subject category), enabling direct integration into model evaluation pipelines without custom parsing or normalization logic.

Unique: Combines breadth (57 academic subjects) with depth (439K questions) and expert curation, making it the largest expert-annotated multiple-choice benchmark at the time of creation. Distributed via HuggingFace's standardized datasets infrastructure with Parquet serialization, enabling zero-copy loading into Pandas/Polars/PyArrow without custom ETL.

vs alternatives: Broader subject coverage and larger scale than earlier QA benchmarks (SQuAD, RACE) while maintaining expert annotation quality, and more rigorous than web-scraped datasets due to academic source validation

Provides pre-split train/validation/test partitions stratified by academic subject, ensuring each subject is represented proportionally across splits. This prevents data leakage where models might memorize subject-specific patterns in training data and enables fair cross-subject generalization testing. The splits are deterministic and reproducible across runs via fixed random seeds.

Unique: Implements subject-stratified splitting at dataset creation time rather than leaving it to users, guaranteeing proportional subject representation across train/val/test without requiring custom sampling logic. This is embedded in the HuggingFace dataset schema rather than requiring post-hoc processing.

vs alternatives: Prevents common evaluation mistakes (subject leakage, imbalanced splits) that plague ad-hoc dataset partitioning, while maintaining simplicity through pre-computed splits

Enables systematic evaluation of language models under zero-shot (no examples) and few-shot (1-5 examples per subject) settings by providing standardized question formatting and answer extraction patterns. The dataset structure supports templating different prompt formats (chain-of-thought, direct answer, explanation-first) while maintaining consistent answer key matching for automated scoring.

Unique: Dataset structure (question + options + answer key) naturally supports both zero-shot and few-shot evaluation without modification, and the subject stratification enables per-subject few-shot analysis to measure learning curves. No proprietary evaluation harness required — standard Python can implement evaluation.

vs alternatives: Simpler and more transparent than closed-source benchmark APIs (e.g., OpenAI Evals) while providing equivalent rigor through expert curation and standardized splits

Enables measurement of how well models trained or evaluated on one set of subjects transfer to held-out subjects, by providing explicit subject labels for every question. This supports leave-one-subject-out evaluation, subject-pair transfer analysis, and domain adaptation studies. The 57-subject taxonomy allows fine-grained analysis of which subject pairs have high transfer (e.g., physics→engineering) versus low transfer (e.g., law→medicine).

Unique: 57-subject taxonomy with balanced representation enables systematic transfer analysis at scale. Subject labels are explicit in dataset schema, eliminating need for post-hoc categorization. The breadth of subjects (STEM, humanities, social sciences, professional) supports analysis of very different domain pairs.

vs alternatives: Larger subject diversity than domain-specific benchmarks (e.g., SciQ for science only) while maintaining expert curation, enabling transfer analysis across truly different knowledge domains

Provides access to the same dataset through multiple Python libraries (HuggingFace datasets, Pandas, Polars, MLCroissant) and serialization formats (Parquet, CSV, JSON), enabling integration into diverse ML workflows without format conversion. Each library interface exposes the same underlying schema (question, choices, answer, subject) but with library-specific optimizations (e.g., Polars for lazy evaluation, Pandas for exploratory analysis).

Unique: Single dataset published simultaneously across multiple library ecosystems (HuggingFace, Pandas, Polars, MLCroissant) with guaranteed schema consistency, rather than maintaining separate dataset versions. Parquet as native format enables zero-copy loading in multiple libraries without conversion.

vs alternatives: More flexible than library-specific datasets (e.g., TensorFlow Datasets) while maintaining consistency better than manual CSV/JSON distribution

Provides explicit categorization of all 439K questions into 57 academic subjects (e.g., abstract_algebra, anatomy, astronomy, business_ethics, clinical_knowledge, etc.) with consistent labeling. This enables filtering, stratification, and analysis at subject level without requiring external knowledge graphs or manual categorization. Subjects span STEM (physics, chemistry, biology), humanities (history, philosophy, literature), social sciences (economics, psychology, sociology), and professional domains (law, medicine, business).

Unique: Explicit subject labels for every question enable filtering without external knowledge graphs or NLP-based categorization. 57-subject taxonomy is comprehensive and expert-validated, covering STEM, humanities, social sciences, and professional domains in single dataset.

vs alternatives: More granular than generic QA datasets (SQuAD, RACE) while maintaining simplicity of flat taxonomy versus complex hierarchical ontologies

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