What do you mean by post‑training in machine learning?

Post‑training refers to the steps you take after the initial training of a machine learning model. It can include fine‑tuning on a domain‑specific dataset, calibrating probability outputs, applying knowledge distillation, or performing model compression. In the context of large language models, post‑training often means taking a pre‑trained transformer and continuing training on a smaller, task‑specific corpus to adapt it to a particular domain. It can also involve adding adapters or prompt‑tuning layers so the base model remains unchanged. Post‑training is essential for improving performance on downstream tasks, reducing overfitting, and ensuring the model’s predictions are calibrated. For example, after training a BERT model on general text, you might fine‑tune it on medical literature to get better results on clinical question answering. Post‑training can also include pruning or quantization to deploy the model on edge devices. The key idea is that the model’s core knowledge is preserved while its behavior is refined for a specific use case.