TOFU Phi 1%: leaderboard
TOFU unlearning test (CMU, 2024): Phi-1.5 tuned on 4,000 QA pairs about 200 fictitious authors must forget the 1% split (2 authors) and keep the rest; forget quality times model utility.
Metric: Forget Quality x Model Utility. Source: huggingface.co. 6 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Phi - Retain Model (WD=0.01) | 0.52 |
| 2 | Phi - Finetune Model (WD=0.01) | 0 |
| 3 | Phi - Grad. Ascent (WD=0.01) | 0 |
| 4 | Phi - Grad. Diff. (WD=0.01) | 0 |
| 5 | Phi - Pref. Opt. (WD=0.01) | 0 |
| 6 | Phi - KL Min. (WD=0.01) | 0 |
Interactive version: theaggregate.ai/benchmark?slug=tofu-phi-1 · How It Works · Data refreshed daily, snapshot 2026-09-05.