TOFU LLaMA 5%: leaderboard

TOFU unlearning test (CMU, 2024): Llama-2-7B-chat tuned on 4,000 QA pairs about 200 fictitious authors must forget the 5% split (10 authors) and keep the rest; forget quality times model utility.

Metric: Forget Quality x Model Utility. Source: huggingface.co. 4 models tracked.

Top models

#ModelScore
1LLaMA 2-7B - Retain Model (WD=0.01)0.6
2LLaMA 2-7B - Grad. Diff. (WD=0.01)0
3LLaMA 2-7B - Grad. Ascent (WD=0.01)0
4LLaMA 2-7B - KL Min. (WD=0.01)0

Interactive version: theaggregate.ai/benchmark?slug=tofu-llama-5 · How It Works · Data refreshed daily, snapshot 2026-09-05.