MathEd-PII: leaderboard
Metric: F1 (%) of PII span detection in the MathEd-PII benchmark: 1,000 U.S. math tutoring transcripts (115,620 messages) with 1,995 PII spans of 17 types replaced by realistic surrogates; each message is sent with the instruction to return every PII span and its type as JSON; F1 of exact span matches against the gold surrogates; the basic prompt (a PII specialist listing the 17 types); higher is better. Source: arxiv.org. Saturation forecast: Around January 2027. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Pro | 81.8 | #77 |
| 2 | Claude Opus 4.5 | 81 | #79 |
| 3 | Gemini 2.5 Flash | 80.7 | #237 |
| 4 | GPT-5.2 | 73.2 | #105 |
| 5 | Claude Haiku 4.5 | 71 | #271 |
| 6 | GPT-4 Turbo | 64.6 | #389 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=mathed-pii · How It Works · Data refreshed daily, snapshot 2026-10-11.