created_at string | database_count int64 | total_rows int64 | total_index_rows int64 | databases list |
|---|---|---|---|---|
2026-07-25T14:35:06.450107+00:00 | 48 | 1,295,400 | 1,288,288 | [
{
"dataset": "noise_025",
"repo": "astropy_astropy",
"json": "/data/dyx/project/openhands/swe-bench-tool/commit_noise_experiment/mixed_json/noise_025/astropy_astropy_commits.json",
"json_sha256": "5a70ae82e8f1211f6171cacadf5372f794e3c77c808751b12c9e5d0675c6d6e3",
"row_count": 47643,
"index_c... |
OpenHands Commit Noise Databases
This dataset contains commit-retrieval databases for 12 SWE-bench repositories at five noise ratios: 0%, 25%, 50%, 75%, and 100%.
Each archive expands to noise_NNN/<repository>/ directories containing:
commits.db: SQLite commit recordscommits.faiss: normalized inner-product FAISS indexcommits.meta.jsonl: FAISS row-to-commit metadatacommits.index_meta.json: embedding and index configuration
The 0% archive is an exact file-level copy of the original database. The noisy datasets use deterministic nested selections with seed 42, so lower-ratio selections are subsets of higher-ratio selections. Of 644,099 planned noisy-field values, 175 generation gaps use the unchanged source field and are recorded in source_fallbacks.json.
The four generated ratios contain 323,850 commit rows each, including 322,072 indexed rows with a non-empty problem description. Embeddings use BAAI/bge-large-en-v1.5 with dimension 1024, normalization enabled, and inner-product search.
Use SHA256SUMS to verify downloads. database_validation_manifest.json records exhaustive SQLite/JSON equality checks, FAISS counts, metadata checks, and database file hashes for the generated ratios. noise_000_snapshot.json records hashes for the original 0% files.
Extract an archive with:
tar --zstd -xf commit_databases_noise_000.tar.zst
- Downloads last month
- 29