datasets-lint

ruff for training data. Audit a dataset for duplicates, PII, distribution problems, label conflicts, schema violations and eval-set contamination — then fix it — offline, on CPU, in one command, with CI-friendly exit codes.

$ uvx datasets-lint run train.jsonl --report out.html
  ✓ stats           120,000 rows · 0.00% empty · worst column 0.0% null
  ✗ dup             18.20% near-duplicate (4,211 clusters), 12.00% exact
  ✗ pii             0.41% of rows contain PII (email 487, phone 12, api_key 1)
  ✗ contamination   3.10% of rows overlap an eval set (gsm8k 3.10%)
  ✓ schema          schema ok (2 columns)
  ✓ leakage         no label conflicts

  verdict: error — 3 errors, 0 warnings, 3 ok · 13.43s
  report → out.html

Install

uvx datasets-lint --help        # run without installing
pip install datasets-lint

The 30-second tour

datasets-lint run data.jsonl                                  # audit
datasets-lint run data.jsonl --fix data.clean.jsonl           # codemod: dedupe, drop empties, redact PII
datasets-lint run data.jsonl --profile strict --report out.html
datasets-lint run data.jsonl --sarif out.sarif                # GitHub code scanning
datasets-lint run data.jsonl --baseline main.json             # did this PR make it worse?
datasets-lint checks                                          # list checks

Exit codes: 0 ok · 1 threshold breached · 2 usage error.

Checks

Full reference: docs/checks.md.

Codemods (--fix)

Every repair is planned first, counted, and written to a new file — the input is never touched.

$ datasets-lint run train.jsonl --fix train.clean.jsonl
  fixes: 327 rows dropped (dup 326 · stats 1) · 26 redactions (email 13 · phone 13)
  fixed → train.clean.jsonl

Details and recipes: docs/codemods.md.

Profiles, suppressions, config

Start with datasets-lint init train.jsonl, then review the inferred columns and percentage thresholds. Existing configs require --force to overwrite. See configuration for the sampling and margin rules.

[defaults]
profile = "strict"        # default | strict | ci | dev
sample = 100000
seed = 0
fail_on = "error"

[checks.dup]
max_near_dup_pct = 2.0

[[suppress]]              # auditable, expiring exceptions
check = "pii"
rows = [12, 13]
reason = "public support addresses, DATA-118"
expires = "2027-01-01"

Full reference: docs/configuration.md.

CI & Marketplace

The Action is listed on the GitHub Marketplace (categories: code-quality, testing).

- uses: horizonbymuneeb/datasets-lint@v1.1.0
  with:
    paths: data/train.jsonl data/test.jsonl
    fail-on: error
- uses: github/codeql-action/upload-sarif@v3
  with:
    sarif_file: datasets-lint.sarif

- uses: horizonbymuneeb/datasets-lint@v1.1.0
  with:
    paths: data/train.jsonl data/test.jsonl
    fail-on: error
- uses: github/codeql-action/upload-sarif@v3
  with:
    sarif_file: datasets-lint.sarif

Pre-commit, Docker, baseline diffs: docs/integrations.md.

Programmatic API

from datasets_lint import audit, fix, Config

report = audit("train.jsonl", config=Config(profile="strict"))
print(report.worst.value)  # "ok" | "warn" | "error"

stats = fix("train.jsonl", "train.clean.jsonl", only=["dup", "pii"])
print(stats["dropped_by_fixer"], stats["redactions"])

Plugins

Third-party checks are normal packages with an entry point — no fork, no vendoring. See docs/plugins.md.

Numbers

Measured on 4 vCPU / 8 GB, no GPU (a laptop-class cloud VM):

Deterministic: same file + config + seed ⇒ identical JSON report.

Contributing

New checks are the best way to contribute: one file, one fixture, one test. See CONTRIBUTING.md and the good first issue list.

License

MIT. Bundled eval data: GSM8K, HumanEval and MMLU, all MIT — see src/datasets_lint/data/eval/manifest.json for sources and checksums.