Post-training (RLHF, RLVR and after)
Where a raw predictor becomes an assistant — and, lately, a reasoner.
Anda sedang melihat entri penjelas konsep: apa arti istilah ini dalam bahasa yang mudah dipahami, dan entri-entri yang dikaitkannya.
After pre-training, models are shaped: supervised fine-tuning on demonstrations, RLHF against human preferences, and now reinforcement learning on verifiable rewards (math checkers, passing tests) that produced the reasoning models. Post-training's share of total compute is rising sharply — labs describe it as the new frontier of differentiation, which is why data-labeling firms and RL environments became strategic assets.
Bagaimana ini terhubung ke bagian lain dari stack
Tautan catatan ini sekilas — klik grup untuk langsung menuju ke sana.
Sumber
DeepSeek-R1 (RL for reasoning) ↗
Rekaman ditinjau per 2026-01-15 — lihat Metodologi.