[Leaderboard] // SFS
Selective Forgetting Score
The headline score — forgetting done right
How well a model forgets a target in plain conversation while staying useful. SFS combines Forget Quality and Utility so it can't be gamed: a model that refuses everything scores zero, and a model that forgets nothing scores zero too. Higher is better.
← All leaderboards#ModelScore
1GPT 5.589.3[84.2, 93.9]
2Moonshot Kimi K2.7 Code85.8[79.3, 92.0]
3GLM-5.283.6[76.9, 88.6]
4Qwen 3.6 Plus83.4[78.1, 87.8]
5DeepSeek V4 Pro82.4[74.7, 88.7]
6Gemini Flash 3.5 (preview)78.7[71.3, 85.5]
7GLM-5.178.5[71.3, 85.1]
8Claude Fable 578.2[70.6, 84.3]
9LLaMa 3.3 70B Instruct73.3[65.1, 80.7]
10Gemma 12B IT72.8[66.1, 78.6]
11Claude Opus 4.872.6[65.0, 79.3]
12Claude Opus 4.772.5[66.1, 79.0]
13Grok 4.2065.5[57.2, 72.4]
14Qwen3 Coder Plus64.7[57.7, 70.9]
15Gemma 12B IT Obliterated61.5[54.5, 68.0]
TL;DR
- The single best number for comparing models on instructed forgetting.
- Blends two opposing goals — suppress the target, keep everything else working.
- Immune to the cheap tricks: blanket refusal and blanket recall both score 0.