PrometheanStudio/promethean-rust-corpus

Promethean Rust Corpus V2

A curated Rust-focused corpus assembled from public source repositories, documentation, and programming exercises.

Current statistics

• Verified unique rows in merged Parquet: 82,645
• New Rust samples added in expansion v3: 31,003
• New Rust source lines reported in expansion v3: 12,281,875
• Original-release repository candidates: 113
• Original-release repositories completed: 103
• Original-release repositories skipped for license review: 9

Sample types

Type	Unique rows
code	45,910
docs	5,504
exercise	228

Largest contributing repositories

Repository	Extracted samples before global deduplication
astral-sh/ruff	1,400
rust-lang/cargo	1,400
wasm-bindgen/wasm-bindgen	1,400
embassy-rs/embassy	1,371
SeaQL/sea-orm	1,325
DioxusLabs/dioxus	1,324
leptos-rs/leptos	1,320
esp-rs/esp-hal	1,319
emilk/egui	1,293
rust-lang/rust-analyzer	1,281
nushell/nushell	1,279
gfx-rs/wgpu	1,276
rust-lang/rust-clippy	1,273
tokio-rs/tokio	1,269
zellij-org/zellij	1,257

Dataset structure

The Parquet dataset contains:

• text: the source code or documentation sample
• sample_type: code, docs, exercise, or mixed
• source_repo: primary source repository
• source_file: primary source path
• license: detected source license information
• content_sha256: content hash used for exact deduplication
• source_count: number of source records associated with the row
• source_records_json: repository, path, and license provenance

Additional files:

• repository_report.csv: extraction status and sample counts
• source_manifest.jsonl: source provenance for each unique sample
• dataset_metadata.json: machine-readable build statistics

Filtering and limitations

Repositories are checked against a conservative list of recognized permissive licenses. Repositories with unknown or ambiguous licenses are skipped for manual review.

Exact duplicate sample content is merged while preserving source provenance. Large repositories are capped to reduce source imbalance. Some generated files, oversized files, and samples matching common secret patterns are excluded.

The license field is repository-level metadata, not a legal determination that every file is covered by the same license. Individual files may have different terms. Review source licenses and attribution requirements before redistribution or commercial use.

This corpus is intended for research and model training. Its existence does not establish that training on every included sample is appropriate for every use.

Reproducibility

The repository candidate list, extraction settings, per-repository reports, content hashes, and source manifest should be retained with each release.