Large model training in 2023 is a simple two-step process: Pretraining is to throw massive data (books, web pages, codes, etc.) to the model so that it can learn the basic laws of language.
Post-training is then polished by aligning with human preferences (such as dialogue, question and answer).
A common method is RLHF (reinforcement learning with human feedback).
馃憠 It鈥檚 like: learn basic textbooks first (pre-training), and then correct mistakes through teacher guidance and exams (post-training).
The large model training in 2025 has been upgraded to "four processes": pre-training is more subdivided, basic data-mixture (Base data-mixture), more code data (More code data-mixture), plus synthetic data (synthetic data-mixture) 馃憠 The model is not only "reading", but also working hard to answer questions and do simulation questions.
Midtraining (Context-expansion): The model can read longer content, such as hundreds of thousands of words, without losing information.
Reasoning heavy: better at solving problems and analyzing steps.
馃憠 Similar to practicing logic and endurance in a high-intensity training camp.
Post-training is more specialized SFT (supervised fine-tuning) DPO/RL (preference optimization, reinforcement learning) 馃憠 Not just "correction", but more refined carving.
Model merging combines models with different specialties, such as forming the "strongest team".