TOFU LLaMA 1%

TOFU evaluates machine unlearning for large language models on fictitious author QA data; this leaderboard variant reports LLaMA submissions at the 1% forget-set setting.

22rows
product_scoreprimary metric
2026-05-06sampled

Metadata

ID
tofu_llama_1p
Category
Safety
Release
2024-01-11

Metrics

Product Score, Model Utility, Forget Quality, Prob. Retain, ROUGE Retain, Truth Ratio Retain, Prob. Real Authors, ROUGE Real Authors, Truth Ratio Real Authors, Prob. Real World, ROUGE Real World, Truth Ratio Real World, Prob. Forget, ROUGE Forget, Truth Ratio Forget

Latest Results

Rank Subject Product Score Model Match Provenance Sampled
1 Retain Model (WD = 0.01) 0.62 — Imported 2026-05-06
2 Pref. Opt. (WD = 0.01) 0.01 — Imported 2026-05-06
3 Grad. Ascent (WD = 0.01) 0.01 — Imported 2026-05-06
4 KL Min. (WD = 0.01) 0.01 — Imported 2026-05-06
5 Pref. Opt. (WD = 0.01) 0.00 — Imported 2026-05-06
6 KL Min. (WD = 0.01) 0.00 — Imported 2026-05-06
7 Grad. Ascent (WD = 0.01) 0.00 — Imported 2026-05-06
8 Grad. Diff. (WD = 0.01) 0.00 — Imported 2026-05-06
9 Grad. Diff. (WD = 0.01) 0.00 — Imported 2026-05-06
10 Grad. Ascent (WD = 0.01) 0.00 — Imported 2026-05-06
11 Grad. Diff. (WD = 0.01) 0.00 — Imported 2026-05-06
12 KL Min. (WD = 0.01) 0.00 — Imported 2026-05-06
13 Pref. Opt. (WD = 0.01) 0.00 — Imported 2026-05-06
14 Pref. Opt. (WD = 0.01) 0.00 — Imported 2026-05-06
15 Grad. Diff. (WD = 0.01) 0.00 — Imported 2026-05-06
16 Finetune Model (WD = 0.01) 0.00 — Imported 2026-05-06
17 Pref. Opt. (WD = 0.01) 0.00 — Imported 2026-05-06
18 Grad. Diff. (WD = 0.01) 0.00 — Imported 2026-05-06
19 KL Min. (WD = 0.01) 0.00 — Imported 2026-05-06
20 Grad. Ascent (WD = 0.01) 0.00 — Imported 2026-05-06
21 KL Min. (WD = 0.01) 0.00 — Imported 2026-05-06
22 Grad. Ascent (WD = 0.01) 0.00 — Imported 2026-05-06