Datasets:
Sources and file access
The data comes from GitHub search and repository API responses, commit change responses and patch files. The four tables expose commit messages, identities, repository observations and code changes directly.
Repository-scoped commit identities use numeric repository IDs when unambiguous. Recorded names remain visible. Source observations with the same identity contribute paired agent/channel labels and available metadata. Disagreeing nonempty metadata values remain null and are named in metadata_conflict_fields.
Commit and repository source references are under indexes/. To inspect an original response, locate its source_artifact_id in indexes/artifacts/. The member_manifest_partition identifies the compressed JSONL member inventory, which gives archive paths, member names and fragment offsets. Concatenate fragments in offset order and verify the original byte count and SHA256. Original record indexes locate the JSONL line or API array item within the recovered file.
Patch text is stored in data/patches/. Encoding patch as UTF-8 reproduces the bytes; bytes and sha256 verify that content. Patch IDs refer to original files, so distinct files remain separate even when their content hashes match.
SHA256SUMS lists delivery-file hashes. manifests/archives.jsonl describes the retained original archives; manifests/release.json records the table sizes and source-file inventory. These files support download verification and source inspection.