Datasets:

---

piimb

/

pii-masking-benchmark-results

like 0

Follow

PII Masking Benchmark 2

Peer review request for the AANA + Presidio submission

This PR is intended as a benchmark-reviewable submission for mindbomber/aana-presidio-piimb-policy-v1.

What is being evaluated:

• Base detector: microsoft/presidio-analyzer predefined recognizers.
• AANA-enhanced system: Presidio outputs plus explicit verifier/correction layers for offset validity, PII-pattern evidence, overlap correction, and high-risk identifier coverage.
• This is a zero-parameter policy architecture paired with a specialist detector, not a trained transformer checkpoint.

Full PIIMB run:

• Dataset: piimb/pii-masking-benchmark
• Dataset revision: df8299e90ff053fa6fd1d3678f6693a454f4ecc0
• Subset: sentences
• Metric/schema: PIIMB 0.2.0

Ablation result:

• Presidio-only average masking F2: 0.4492985573
• Presidio + AANA average masking F2: 0.5629171363
• Absolute F2 gain: +0.1136185790
• Presidio-only average recall: 0.4008557794
• Presidio + AANA average recall: 0.5159532273
• Absolute recall gain: +0.1150974479

Per-source AANA F2:

• OpenPII: 0.4879480402
• Gretel: 0.6281397502
• Nemotron-PII: 0.6161414756
• Privy: 0.5194392792

Repro/model card:
https://huggingface.co/mindbomber/aana-presidio-piimb-policy-v1

Scope and limitations:

• The claim is only that the AANA verifier/correction layer improved masking F2 and recall over the Presidio-only baseline on this PIIMB run.
• I am not claiming state-of-the-art performance, guaranteed PII removal, or production readiness for regulated workflows.
• Lower strict/type NER scores are expected because this architecture optimizes broad masking coverage and offset correctness rather than exact source-dataset entity taxonomy.

The submission was also checked through an AANA model-evaluation release gate before publication: gate_decision=pass, recommended_action=accept, candidate_gate=pass, no hard blockers. Review on methodology, result-file shape, and whether this policy-architecture submission fits leaderboard criteria would be appreciated.

Public Try AANA endpoint

The interactive AANA demo Space is now live and should be used as the public
"try AANA" endpoint for this submission and future benchmark references:

https://huggingface.co/spaces/mindbomber/aana-demo

It accepts:

• candidate answer/action
• evidence
• constraints

It returns:

• AANA route: accept, revise, ask, defer, or refuse
• AIx score
• hard blockers
• suggested revision/route
• JSON audit summary

Canonical AANA model card:
https://huggingface.co/mindbomber/aana

Benchmark context remains bounded to this PIIMB ablation: Presidio-only average
masking F2 0.4492985573; Presidio + AANA average masking F2 0.5629171363;
absolute F2 gain +0.1136185790. This is not a claim of state-of-the-art
performance, guaranteed PII removal, guaranteed hallucination removal, trained
weights, or production readiness.

Same comment as https://huggingface.co/datasets/piimb/pii-masking-benchmark-results/discussions/2 : I'm not looking to add regex baselines to the benchmark at this point.