---

tags:

• ogolne

---

Ml post‐training

post-training

It is the art and scienceof transforming foundation models into useful, reliable, safe, and efficient production systems. It encompasses everything that happens after the foundation model itself has been created:

• fine-tuning on domain-specific data,
• alignment with human values,
• optimization for deployment,
• evaluation in real-world scenarios, and
• ensuring safety and compliance.

post-training refers to any process where gradients flow through the model after pre-training, such as

• supervised fine-tuning (SFT),
• reinforcement learning from human feedback (RLHF), and
• preference optimization.