Removing RLHF Protections in GPT-4 via Fine-Tuning

Keyphrases

• Reinforcement Learning 100%
• Human Feedback 100%
• GPT-4 100%
• Large Language Models 80%
• Weak Models 40%
• Success Rate 20%
• In Tandem 20%
• Training Data 20%
• Training Examples 20%
• Dual Use 20%
• Attacker 20%
• Fine-tuning Strategy 20%

Computer Science

• Reinforcement Learning 100%
• Generative Pre-Trained Transformer 4 100%
• Large Language Model 80%
• Training Data 20%
• Attackers 20%
• Training Example 20%