Fine-tuning

Continuing to train an already-trained model on a small amount of new-task data, so it gets better at that specific job.

Fine-tuning is the most common form of transfer learning: instead of starting from random parameters, you start from a pretrained model and keep training it on data from the target task, reusing the general knowledge the model already learned, which saves both data and compute. By how much gets updated, full fine-tuning changes every weight; alternatively, most layers can be frozen and only a subset trained, or a parameter-efficient method like LoRA (Low-Rank Adaptation, which inserts a small number of trainable low-rank matrices) can be used to train only a tiny fraction of the parameters. In embodied AI, the most common way to get started with an open-source VLA is to fine-tune it on demonstration data collected on your own robot. Too little data, or training for too long, easily leads to overfitting, and can also make the model forget capabilities it previously had — a problem called catastrophic forgetting.

ExampleThe openpi documentation gives this reference: fine-tuning π0 with LoRA needs at least 22.5GB of GPU memory (such as an RTX 4090), while full-parameter fine-tuning needs at least 70GB (such as an A100 or H100).

Also called

Finetuning

Related

Pre-training · Post-training · Full Fine-Tuning · LoRA · Supervised Fine-Tuning · Catastrophic Forgetting

Sources

Wikipedia: Fine-tuning (deep learning)
Google Machine Learning Glossary: fine-tuning
GitHub: Physical-Intelligence/openpi