Manju46/swift-test-pairs

SwiftTestPairs

20,801 Swift source↔unit-test pairs ("focal method" pairs) mined from permissively licensed public GitHub code.

Applies the methods2test (MSR 2022) methodology to Swift, which had no equivalent dataset.

⚠️ This dataset contains NO source code

Rows are references + derived metadata: repository, file paths, content MD5s, and measured properties. Reconstruct file contents from the two public upstream datasets with resolve.py (⚠️ streams ~4 GB from the Hugging Face Hub; use --tier 2 for the 569-pair benchmark subset). This mirrors what Stack-Edu does upstream — it avoids redistributing code, and it honours upstream opt-outs automatically: a file withdrawn from the source corpora simply fails to resolve.

The novel contribution here is the pairing, not the source.

Schema

column	description
pair_id	md5 of corpus:repo:source_path:test_path
source_corpus	iva or stackedu — which upstream dataset the row resolves against
repo_name, source_path, test_path	location of the focal file and its test
target_type	the type under test, derived from the test filename
source_content_md5, test_content_md5	verify resolution fidelity
source_bytes, test_bytes	file sizes
n_assertions, n_test_methods	measured from the test file
source_score, test_score	Stack-Edu educational-quality score (null for iva)
license	SPDX name (iva) or bucket label (stackedu)
test_uses_swift_testing	imports Swift Testing rather than XCTest
has_legacy_swift	contains Swift 2/3 constructs no current toolchain accepts
join_method	filename or declaration
tier	1 = pair only · 2 = verified to build with Swift 6.4
split	train / valid / test, assigned by repository

Tiers

tier	pairs	repos	median assertions	use
1	20,232	~7,400	11	retrieval, idiom mining, assertion-level comparison
2	569	157	14	benchmarking — repo builds clean under Swift 6.4

Tier 2 repos were each built with swift build --build-tests using their own Package.swift. 47% of eligible repos built; only 31.5% of eligible pairs survived, because larger repos hold more pairs and fail more often.

⚠️ Tier 2's test split holds only 30 pairs. For evaluating a generator no training split is needed — use all 569.

Construction

2 corpora (3.2M files) → permissive only → drop vendored trees (Pods/, Carthage/…)
  → classify TEST / TEST_SUPPORT / SOURCE
  → join: filename convention (22,912) + declaration index (1,045)
  → exact dedup, cross-source (−1,868, 7.8%)
  → MinHash/LSH near-dup, Jaccard 0.8 (−1,288, 5.8%)
  → repo-level 80/10/10 split          = 20,801 pairs

Both dedup passes matter: the two corpora overlap 62.6% at pair level despite sharing only 13.3% of repositories. Repo-level splitting alone does not catch this — the crawls spell repo_name differently for the same project.

Validation

check	tier 1	tier 2
test file mentions target_type	98.1%	97.7%
source file contains target_type	94.3%	92.4%
resolver round-trip, MD5 verified	—	569/569 (100%)

Tier 2 was additionally verified by building every repo with swift build --build-tests.

Known limitations

• Content vintage ≈ 2023. Both upstream corpora derive from ~2023 crawls.
• XCTest, not Swift Testing. Only 126 of 20,801 pairs (0.6%) use Swift Testing.
• 7.2% contain Swift 2/3 syntax (has_legacy_swift) that no current compiler accepts.
• Licenses are repo-level, not file-level. A repo declared MIT can contain vendored third-party code. Treat license as a signal, not a guarantee.
• target_type is heuristic — derived from the test filename, not semantic analysis. 757 pairs (3.6%) have extension-style targets (String+Extensions, Observable+Time) that name a file rather than a type. The source↔test pairing is still correct; only the target_type label is imprecise.
• Tier 1 pairs are not known to compile. Most will not, in isolation: real Swift repos are multi-target with external dependencies.

Provenance

• mvasiliniuc/iva-swift-codeint — BigQuery GitHub dump (12,435 pairs)
• HuggingFaceTB/stack-edu → The Stack v2 → StarCoder2Data → Software Heritage (8,366 pairs)

Per-row licenses are in the license column. Cite upstream, not a mirror.

Removal requests

If your repository appears in this dataset and you would like it removed, please open a Discussion on this dataset and name the repository. I will drop those rows and re-publish. No justification needed.

Note that this dataset contains no source code — only repository names, file paths, content fingerprints and measurements. The code itself remains where it has always been, in the upstream corpora listed above.

Citation

@misc{swamymanju2026swifttestpairs,
  title  = {SwiftTestPairs: Swift source-to-test pairs for test generation research},
  author = {Swamy Manju},
  year   = {2026},
  url    = {https://huggingface.co/datasets/Manju46/swift-test-pairs}
}