Post-training

Post-training adapts an existing model to a defined domain or behavior, but it is not the default starting point. Prompt design, retrieval, permissioned tools, deterministic system controls, and a measured baseline on held-out cases come first. If that baseline passes, it is packaged without changing model weights. If it misses, the team diagnoses the gap before choosing a method.

Diagnose before adapting

A context, retrieval, tool, or process gap calls for a system fix rather than training. A behavior-demonstration gap may justify supervised fine-tuning (SFT) on approved expert examples. A preference gap may justify preference optimization such as direct preference optimization (DPO). Reinforcement learning (RL) or reinforcement fine-tuning (RFT) is conditional: use it only when experts can define a reliable reward, rollouts are controlled, and an independent verifier can detect failure or reward gaming. These are alternative routes, not a mandatory sequence.

The gate is the contract

A candidate returns to held-out evaluation, then relevant safety and adversarial tests, agentic task evaluation, and shadow or operational proof. It is packaged only if every applicable gate passes without unacceptable regression; otherwise the team stops or revises the diagnosed layer. Training and evaluation data remain separated, and provenance, retention, privacy controls, and the permitted training environment are documented in scope.

Who owns the result

Ownership, model-license constraints, deployment rights, and handoff artifacts are written into the proposal. Where the selected model and environment permit it, applicable artifacts can include adapters or weights, evals, data documentation, configuration, and runbooks.