REDACT: leaderboard
REDACT (Vats et al., 2026) tests personal-information detection in synthetic records across 25 languages and 51 entity types, generated under nine controlled variation axes. A detector returns every personal-information span with its type; the board is partial-overlap micro-F1 on a locked, language-stratified sample of 1,000 records (a same-type span with any overlap counts), scaled to 0-100 (higher is better).
Metric: Partial-overlap micro-F1 (x 100; predicted personal-information spans matching a gold span of the same type with any overlap, 51 entity types, locked language-stratified sample of 1,000 records in 25 languages). Source: arxiv.org. Saturation forecast: Around February 2027. 4 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Sonnet 4.6 | 63.6 | #158 |
| 2 | GPT-4.1 | 59.7 | #492 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=redact · How It Works · Data refreshed daily, snapshot 2026-11-12.