Meta: Llama Guard 4 12B vs The Stack v2

Classifies both text and image inputs against a taxonomy of unsafe content categories (violence, sexual content, hate speech, etc.) using a fine-tuned Llama 4 Scout backbone with multimodal encoders. The model processes inputs through separate text and vision pathways, then aggregates representations to produce safety risk scores and category labels. Built on instruction-tuned safety classification patterns established in Llama Guard 3, extended with visual understanding for detecting unsafe imagery.

Unique: First Llama Guard iteration with native multimodal (text + image) safety classification using a unified Llama 4 Scout backbone, rather than separate text-only classifiers or vision models bolted together. Extends instruction-tuned safety taxonomy from Llama Guard 3 with visual understanding for detecting unsafe imagery without requiring separate image classifiers.

vs alternatives: Handles text and image safety in a single model call with shared semantic understanding, whereas alternatives like OpenAI Moderation API (text-only) or separate image classifiers require multiple API calls and lose cross-modal context.

Maps input content to a predefined taxonomy of unsafe categories (violence, sexual content, hate speech, illegal activities, etc.) using instruction-tuned classification. The model was fine-tuned on safety-labeled datasets to recognize nuanced violations within each category, producing granular category-level confidence scores rather than binary safe/unsafe decisions. Supports hierarchical reasoning about content severity across multiple harm dimensions simultaneously.

Unique: Uses instruction-tuned fine-tuning on safety-labeled data to produce multi-dimensional category scores in a single forward pass, rather than training separate binary classifiers per category or using rule-based heuristics. Inherits Llama Guard 3's taxonomy design but extends it with visual understanding.

vs alternatives: Provides granular per-category scores in one API call, enabling policy-based routing, whereas binary classifiers (safe/unsafe) require downstream logic to determine which violation type occurred, and rule-based systems are brittle to paraphrasing.

Applies instruction-following capabilities from the Llama 4 Scout base model to safety classification tasks, enabling the model to understand nuanced safety instructions and apply them consistently. The fine-tuning process teaches the model to reason about context, intent, and harm potential rather than matching keywords. This allows classification of subtle violations (e.g., veiled threats, coded hate speech) that simple pattern matching would miss.

Unique: Leverages instruction-tuned capabilities from Llama 4 Scout to perform contextual reasoning about safety violations, rather than relying on keyword matching or shallow pattern recognition. Fine-tuning teaches the model to understand intent, context, and nuance in safety classification.

vs alternatives: Detects obfuscated or contextually-dependent violations that keyword-based systems miss, and maintains consistency across paraphrases, whereas rule-based classifiers require exhaustive enumeration of violation patterns and fail on novel phrasings.

Exposes safety classification through OpenRouter's API, enabling batch processing of content at scale without managing inference infrastructure. Requests are routed through OpenRouter's load-balanced endpoints, supporting concurrent classification of multiple text/image inputs. The API abstracts away model serving complexity, providing a simple HTTP interface with standard request/response formats.

Unique: Provides managed API access to Llama Guard 4 through OpenRouter's infrastructure, eliminating the need for self-hosted deployment while maintaining multimodal safety classification capabilities. Abstracts model serving, scaling, and versioning complexity behind a simple HTTP interface.

vs alternatives: Eliminates infrastructure management burden compared to self-hosted deployment, and provides built-in scaling/reliability, whereas self-hosting requires GPU procurement, model optimization, and operational overhead.

Processes images through a vision encoder integrated into the Llama 4 Scout backbone to detect unsafe visual content (violence, sexual imagery, hate symbols, etc.). The vision pathway extracts visual features that are then fused with text embeddings for joint classification. This enables detection of unsafe imagery even without accompanying text, and allows the model to understand visual context when classifying text+image pairs together.

Unique: Integrates vision encoding directly into the Llama Guard 4 architecture for end-to-end multimodal safety classification, rather than using separate image classifiers or post-hoc fusion of text and image scores. Enables joint reasoning about image+text pairs with shared semantic understanding.

vs alternatives: Classifies images and text together in a single model with shared context, whereas separate classifiers (e.g., CLIP for images + text classifier) require multiple API calls and lose cross-modal reasoning about hateful memes or context-dependent visual harms.

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