🖼️ 인포그래픽

🖼️ 4컷 인포그래픽

💡 한 줄 요약

Teaching models to forget: Selective unlearning with Amazon Nova | Artificial Intelligence Skip to Main Content Artificial Intelligence Teaching models to…

📌 핵심 요약

• Teaching models to forget: Selective unlearning with Amazon Nova | Artificial Intelligence Skip to Main Content Artificial Intelligence Teaching models to forget: Selective…
• Amazon SageMaker AI supports DPO training with both full-rank and LoRA approaches for Amazon Nova and over 20 open-weight models , and Amazon SageMaker HyperPod provides…
• The LoRA-based approach is particularly well-suited to this problem: it isolates unlearning to a small set of adapter parameters, preserves the base model’s integrity, and enables…
• For researchers and practitioners interested in further exploring preference-based unlearning, Amazon SageMaker AI offers DPO training recipes that can serve as a foundation for…
• To verify this, we evaluate the customized model on three utility benchmarks: instruction following, mathematical reasoning (Math Mini), and code generation (MBXP Python): Model…

🧩 주요 포인트

1. Teaching models to forget: Selective unlearning with Amazon Nova | Artificial Intelligence Skip to Main Content Artificial Intelligence Teaching models to forget: Selective…
2. Amazon SageMaker AI supports DPO training with both full-rank and LoRA approaches for Amazon Nova and over 20 open-weight models , and Amazon SageMaker HyperPod provides…
3. The LoRA-based approach is particularly well-suited to this problem: it isolates unlearning to a small set of adapter parameters, preserves the base model’s integrity, and enables…
4. For researchers and practitioners interested in further exploring preference-based unlearning, Amazon SageMaker AI offers DPO training recipes that can serve as a foundation for…
5. To verify this, we evaluate the customized model on three utility benchmarks: instruction following, mathematical reasoning (Math Mini), and code generation (MBXP Python): Model…

🧠 상세 정리

1. 에너지·칩·모델의 연결

Teaching models to forget: Selective unlearning with Amazon Nova | Artificial Intelligence

2. 에너지·칩·모델의 연결

Organizations deploying foundation models (FMs) often encounter a common challenge: model safeguards designed for content moderation can also prevent legitimate, business-critical use cases. A media company summarizing scripts with mature language, a cyber security firm…

3. 에너지·칩·모델의 연결

Because the model learns these safeguards during post-training alignment, prompt engineering alone cannot overcome them. The model’s tendency to deflect is embedded in its parameters, requiring a targeted modification at the model level to selectively adjust this behavior. In…

4. 산업적 의미

Amazon Nova Customizable Content Moderation Settings (CCMS) addresses this by letting approved customers selectively adjust safeguards across four responsible AI (RAI) pillars. These pillars encompass:

5. 적용 사례와 파급효과

Amazon Nova enforces essential, non-configurable controls for responsible use of AI, such as controls to prevent harm to children and preserve privacy.

6. 에너지·칩·모델의 연결

The science behind CCMS is unlearning , a technique for selectively removing learned behaviors from a model’s parameters without retraining from scratch. We train Low-Rank Adaptation (LoRA) adapters to reverse the model’s alignment to specific policies. The result is a custom…

🧾 핵심 주장 / 시사점

• Our evaluation on Amazon Nova 2 Lite demonstrates that rDPO-trained LoRA adapters reduce deflection rates by up to 54 percentage points across RAI policy…
• These techniques power Amazon Nova CCMS, which provides customers with pre-trained adapters ready to deploy using Amazon Bedrock. For researchers and…
• The contributors to this project include Ekraam Sabir, Weitong Ruan, Payal Motwani, Rahul Gupta, Claire O’Brien Rajkumar, Dhwanil Desai, and Nikhil Sanil.

✅ 액션 아이템

• Amazon SageMaker에서 DPO full-rank와 LoRA를 동일 조건으로 준비해 Amazon Nova의 선택적 망각 실험군을 분리한다.
• LoRA 기반 실험은 어댑터 파라미터 집합만 갱신해 기본 모델 무결성 보존을 점검하고, 필요 시 적용 범위를 축소한다.
• 커스터마이즈 모델은 instruction following, Math Mini, MBXP Python 세 유틸리티 벤치마크로 성능과 회복 정도를 정량 점검한다.

❓ 열린 질문

• LoRA 기반 DPO의 어댑터 파라미터 제한이 실제 망각 시나리오에서 기대한 삭제 효과를 충분히 달성하는가?
• Amazon SageMaker AI가 제공한 DPO 레시피를 20개 이상 오픈웨이트 모델에서 적용할 때 어떤 지표가 성공 조건을 판별하는 기준이 될 것인가?
• instruction following·Math Mini·MBXP Python 중 어느 항목에서 성능이 먼저 흔들릴 경우 선호 기반 망각 강도를 어디까지 조정해야 하는가?