Which LLM fine-tuning method should we focus on: instruction tuning, RLHF, or adapter-based tuning?

24 Tem 22:00

We're planning our next LLM project and need to decide which fine-tuning paradigm to invest time in. Option 1: instruction tuning – using a curated set of prompts to teach the model desired behavior. Option 2: reinforcement learning from human feedback (RLHF) – aligning outputs with human preferences via reward models. Option 3: adapter-based tuning – adding small trainable modules while keeping the base model frozen. Which approach do you think gives the best balance of performance and compute cost? Share your reasoning!