label_cleaner

Modular rewrite of the original experiment codebase.

Architecture

Foldered Modules (primary)

• core/
  • models.py — typed dataclasses (PreparedSplit, NoiseBundle, MethodCurves)
  • prep.py — dataset split preparation (prepare_fixed_split)
• data/
  • datasets.py — dataset loading + metadata (DatasetInfo, load_dataset)
• methods/
  • noise.py — noise injectors (inject_outlier, inject_rnd_label, inject_nnar, inject_mnar)
  • pipelines.py — pipeline factories (make_pipeline_a, make_pipeline_b)
  • cleaning.py — methodology functions (clean_datascope, clean_random, clean_cleanlab)
• orchestration/
  • experiments.py — run functions per noise type
  • catalog.py — experiment/pipeline registry
• services/
  • service.py — microservice-style facade (LabelCleanerService)

Example

from pathlib import Path
from label_cleaner.services.service import LabelCleanerService

svc = LabelCleanerService(datasets_dir=Path("datasets"))
result = svc.run(
    dataset="adult",
    noise_type="rnd_label",
    pipeline_key="p1a",
    noise_level=0.2,
)

print(result["baseline"])          # noisy baseline accuracy
print(result["datascope"][-1])     # accuracy after 100% DataScope cleaning
print(result["cleanlab"][-1])      # accuracy after 100% CleanLab cleaning

Notes

• All methodology functions are isolated and reusable.
• Service API is ready to expose via REST/gRPC later if needed.
• Current implementation preserves behavior from the original scripts.