Vs (SFT & DPO & RLHF)

Description

• SFT: Trains on labeled input-output pairs, but does not directly optimize for human preferences.
• RLHF: Uses a reward model and reinforcement learning to align with preferences, but is complex and unstable.
• DPO: Directly optimizes for preference pairs, offering a simpler and more robust alternative to RLHF.