Metadata-Version: 2.5
Name: maskflow-bench
Version: 0.1.0
Summary: MaskFlow benchmark scoring core: load labelled PII data, score detections under strict/partial-overlap precision-recall-F1, write JSON/Markdown reports. Powers `maskflow bench --my-data`.
License: MIT
Requires-Python: >=3.10
Requires-Dist: maskflow-core<0.9,>=0.8.0
Provides-Extra: dev
Requires-Dist: maskflow-pack-india<0.6,>=0.5.1; extra == 'dev'
Requires-Dist: pytest>=8.0; extra == 'dev'
Description-Content-Type: text/markdown

# maskflow-bench

The scoring core behind MaskFlow's accuracy measurements: load a labelled
JSONL corpus, canonicalize its entity taxonomy, score detections under
strict-span and partial-overlap precision/recall/F1, write `results.json`
and `results.md`.

Most people reach this through `maskflow bench --my-data <path>`
(`maskflow-cli`) — see [`docs/bench.md`](../../docs/bench.md) for the
labelled-data schema and command reference.

This package intentionally ships **one** adapter — MaskFlow itself
(`adapters.maskflow_adapter.MaskflowAdapter`) — with no dependency beyond
`maskflow-core`. The multi-adapter comparison against Presidio,
mask-privacy, a naive-regex baseline, and an LLM judge that produced the
published IndiaPII-Bench tables lives in `bench/indiapii/harness/` (repo
dev tooling, not published), and imports this package for the parts that
don't change per adapter.
