Preferred training approach for next-gen LLMs: supervised fine‑tuning, RLHF, or self‑supervised instruction?

26 Eyl 13:45

I'm planning the next phase of my LLM project and need a clear direction. Between three high‑level strategies—(1) supervised fine‑tuning on domain‑specific data, (2) reinforcement learning from human feedback (RLHF) to shape behavior, and (3) a self‑supervised instruction‑following pre‑training loop—which one would you choose as the primary focus? Briefly share why you favor that approach (e.g., data efficiency, alignment quality, scalability). Your insights will help shape the roadmap. Thanks!