StarCoderData

Processes raw code from The Stack (a 3TB+ dataset) through a multi-stage filtering pipeline that applies near-deduplication heuristics (likely MinHash or similar probabilistic techniques) to identify and remove near-identical code blocks across 86 programming languages. The curation preserves language-specific semantics while reducing redundancy, enabling models trained on this data to learn diverse coding patterns rather than memorizing repetitive boilerplate. Outputs a deduplicated 250GB subset suitable for model pretraining.

Applies automated PII (Personally Identifiable Information) detection and removal across the dataset, scanning for patterns like email addresses, API keys, credentials, and personal names embedded in code comments or strings. Uses regex-based and potentially ML-based classifiers to identify sensitive data, then either redacts or removes affected code samples. This ensures the resulting dataset is safe for public distribution and model training without leaking private information.

Implements heuristic-based quality filtering to exclude low-quality, malformed, or non-functional code samples from the dataset. Likely uses metrics such as: file size thresholds (excluding very small or very large files), syntax validity checks (parsing code to ensure it's well-formed), license filtering (excluding code with restrictive licenses), and potentially code complexity or style metrics. Filters are applied per-language to respect language-specific conventions (e.g., Python indentation rules vs. JavaScript semicolons).

Provides code samples across 86 programming languages with language-aware metadata and tokenization support. Each sample is tagged with its language, enabling downstream models to learn language-specific patterns and syntax. The dataset structure supports efficient loading and batching of code by language, allowing models to train on language-balanced or language-specific subsets. Tokenization is deferred to the model training pipeline, but the dataset preserves raw code to enable flexible tokenizer choices.

Augments raw code samples with GitHub metadata including issue descriptions, commit messages, and code change history. This provides semantic context for code snippets, enabling models to learn the relationship between code changes and their motivations/descriptions. The dataset likely includes paired examples of (code, issue description) or (code change, commit message), enriching the training signal beyond syntax-only learning. Enables training on code-to-text and text-to-code tasks simultaneously.

Provides versioned snapshots of the curated dataset with reproducible train/validation/test splits, enabling researchers to compare results across experiments and publications. Uses deterministic splitting logic (likely based on file hashes or fixed random seeds) to ensure the same code samples appear in the same splits across different downloads. Metadata includes dataset version, curation date, and filtering parameters, enabling reproducibility and ablation studies.

Implements streaming-based data loading via Hugging Face Datasets library, enabling researchers to train on the full 250GB dataset without downloading it entirely upfront. Uses lazy loading and on-the-fly batching to load code samples into memory as needed, reducing storage requirements and enabling training on machines with limited disk space. Supports efficient sampling, shuffling, and filtering operations without materializing the full dataset.

Enables fine-grained control over dataset composition by language, allowing researchers to sample code by language distribution, exclude specific languages, or oversample underrepresented languages. Provides language-stratified sampling to ensure balanced training across languages or language-specific fine-tuning. Metadata includes language distribution statistics, enabling informed decisions about dataset composition.