[Topic]
Fine-tuning & post-training
Supervised fine-tuning, preference optimization, reinforcement learning and the data that feeds them.
GuideIntermediate
The complete guide to LLM post-training in 2026
LLM post-training is everything done to a pretrained model to turn raw next-token prediction into useful, reliable behavior: supervised fine-tuning, preference optimization, reinforcement learning and distillation. For most enterprises the question is not which algorithm is best, but whether a measured training run beats a well-built prompt or retrieval system on their own task.
ComparisonAdvanced
SFT vs DPO vs RL: understanding modern post-training methods
SFT imitates demonstrations, DPO learns from preferred-versus-rejected pairs without a separate reward model, and reinforcement learning optimizes a reward by sampling from the model itself, using PPO with a learned reward model or GRPO with verifiable rewards. Most production recipes run SFT first, then add preference optimization or RL only where a measurable gap and the right signal exist.
GuideIntermediate
Synthetic data: how to generate training data for better models
Synthetic data for fine-tuning is training data written by a model: new prompts, teacher responses, or sampled answers filtered down to the correct ones. It works when anchored to real examples and every record is verified, deduplicated and decontaminated against your evaluation set. It fails when it replaces real data, copies a teacher's style without its accuracy, or breaks a provider's terms.