task_categories:
- text-generation
language:
- code
size_categories:
- n<1K
Joern C++ GitHub 300: source inventory
A research inventory for testing C++ parser and source-CFG tooling. It records 300 distinct public GitHub repositories and 578 selected C++ source files, plus two repositories where selection found no suitable file. Source files are not redistributed: manifest.json records exact repository commits, upstream paths, Git blob hashes, SHA-256 hashes, raw source URLs, and GitHub-reported repository SPDX identifiers. The repository-level SPDX value is discovery metadata, not a file-level licensing determination. Review each upstream project and file before downloading or redistributing it.
This is an unqualified source inventory, not a benchmark leaderboard, a verified CFG oracle, a dataset of compilable translation units, or a claim of standard/version coverage. Source files are selected from public GitHub repository search results stratified by project creation date and a compiler topic query. Rankings and API availability introduce selection bias. The original Ninja, Snappy and double-conversion corpora were excluded. GitHub tree truncation is recorded; the largest-file choice applies only to the returned eligible tree, not necessarily the whole repository. Two entries may point to the same project if upstream projects have copied sources; distinct repositories do not prove independent codebases.
Per repository, the selection favors a non-vendored C++ implementation file and, where different, the largest eligible implementation file between 256 bytes and 200,000 bytes. It is a discovery sample rather than a curated set of representative whole projects. Search queries, unsuccessful selection, exact commits and hashes are preserved in the manifest. GNU/C++ standard, actual build flags, generated headers, dependencies, source ownership, license exceptions and target configuration have not been established. Projects may change or disappear; immutable commit URLs and hashes provide verification when still available.
The local initial diagnostic ran Rust-Joern in strict mode against raw files. Such files often require headers and macro expansion, so rejections do not measure parser quality. A separate standalone Clang C++17 qualification was attempted on all 578 files: 564 were rejected or lacked build context; 8 compiler-accepted files were rejected by Rust-Joern after preprocessing, and 6 produced no functions. Details and limited owner-only diagnostics are in diagnostic-summary.json. Neither that setting nor a successful parse proves the original project's build mode or CFG correctness. Original Joern and trustworthy compiler-backed CFG comparisons are future work.
Selection and run tooling: https://github.com/sjkim1127/joern-c-cpp-corpus . The initial tooling is pinned at commit 27fe5ea in that repository. No source file contents or compiler-generated/preprocessed files are uploaded in this dataset.