[Leaderboard] // FORGET QUALITY
Forget Quality
Does the secret stay forgotten under attack?
How thoroughly the forgotten target is suppressed when we actively try to get it back — through direct questions, rephrasings, indirect approaches, other languages, role-play, and more. Harder tricks count for more. Higher is better.
← All leaderboards#ModelScore
1GPT 5.588.0
2Claude Fable 582.9
3Moonshot Kimi K2.7 Code80.9
4Qwen 3.6 Plus79.8
5Gemini Flash 3.5 (preview)79.7
6GLM-5.275.9
7DeepSeek V4 Pro74.8
8GLM-5.173.9
9Claude Opus 4.771.1
10Claude Opus 4.870.1
11LLaMa 3.3 70B Instruct62.6
12Gemma 12B IT59.8
13Qwen3 Coder Plus58.1
14Grok 4.2050.7
15Gemma 12B IT Obliterated44.4
TL;DR
- Measures resistance to recovery, not just polite compliance.
- Probes range from easy (ask again) to hard (multi-step indirection).
- A model that says “I forgot” but leaks under role-play scores low here.