Defects4C SWE-bench-style export

One instance per repository and fixing commit, joined from the local Defects4C export. all includes projects, legacy projects_v1, and additional cached commits. projects restricts to projects/. Duplicate metadata is retained in bug_metadata_json. Cached-only instances use the first parent as base_commit. The single test split is an organizational choice, not a validated benchmark split. No repository-disjoint training partition is claimed.

Load

from datasets import load_dataset
ds = load_dataset("json", data_files="data/all.jsonl", split="train")
# Once this folder is uploaded to a Hugging Face dataset repository:
# ds = load_dataset("YOUR_NAME/DATASET_NAME", "projects", split="test")

Fields and limitations

Core fields follow the SWE-bench dataset structure. base_commit is the buggy revision; head_commit is the fixed revision. patch contains changes outside the recorded test-file list; test_patch contains changes to that list; full_patch contains both. The test-file list is metadata, not an independently validated exhaustive classification. Cached-only rows have no classification: their patch contains all changes.

patch_source=local_git and patch_is_exact=true mean a binary-capable diff was generated directly between the two Git commits. API patches are reconstructed text hunks and are not guaranteed complete or applyable: mode changes, binary data, truncated hunks, or API file pagination may be absent. Known missing file hunks are recorded in omitted_patch_files and flagged cached_api_partial; other cached patches are cached_api_unverified. Missing patches or base mismatches remain explicit. Empty patches do not imply no fix.

problem_statement uses a neutral repair request and the original diagnostic; it does not insert the fixing commit message or gold code. The original errors may be runtime/test output rather than compiler errors. No upstream issue text is invented. original_errors preserves distinct messages; commit_message and cached repair records are reference material that can leak the solution.

FAIL_TO_PASS and PASS_TO_PASS are JSON-encoded empty lists because test transitions were not measured. test_outcomes_verified is false, and version is empty because package versions were not established. This dataset needs repository-specific environments, test runners, and result parsers before it can be scored by the SWE-bench harness. Build success is not test validation.

build_command is provided for projects/ rows supported by checkout_build.py. Legacy and cached-only rows retain available templates/configuration but have no claimed runnable build command. build_artifacts_json includes existing pre/post build scripts, logs, and summaries; no new builds are performed during export. build_templates_json retains project and shared Jinja templates.

Fields ending in _json, plus original_errors, source_files, test_files, omitted_patch_files and the test-transition fields, are JSON strings; decode them with json.loads. This preserves heterogeneous source schemas in Arrow. Other availability fields are booleans and all remaining columns are strings.

source_data.tar.gz preserves the complete original data/, projects/, and projects_v1/ inputs and reproduction/export scripts, including unjoined records. Cached source snapshots retain their original provenance; no revision identity is asserted beyond the source record. checksums.json contains file SHA-256 sums.

Coverage

{
  "instances": 365,
  "repositories": 44,
  "collections": {
    "projects": 119,
    "projects_v1": 241,
    "cached_only": 5
  },
  "patch_sources": {
    "cached_api_unverified": 295,
    "cached_api_partial": 67,
    "local_git": 2,
    "missing": 1
  },
  "with_original_error": 362,
  "with_build_artifacts": 2,
  "verified_test_outcomes": 0
}

Reproduce

From the original export (or after extracting source_data.tar.gz):

python3 export_hf_dataset.py --output /path/to/new/dataset

The exporter uses the Python standard library and performs no downloads. Put existing builds under build/PROJECT/COMMIT/ before exporting to include their artifacts and derive exact patches from their Git objects. Use a new destination for each export. The archive does not include cloned repositories or binaries.

The JSONL files and this card use the Hugging Face repository layout. Upload this directory as a dataset repository to publish it. No upload is performed by the exporter.

Licensing

This export does not assign a new license to upstream code or cached content. Each project's original license and attribution requirements continue to apply. The source export does not establish a single dataset-wide redistribution license.