Foundation model
A broadly pretrained model used as the starting point for many products, then adapted with prompts, retrieval, or fine-tuning.
See it
What it is
A foundation model is trained broadly enough to become the starting point for many different applications. Instead of training a translator, classifier, and chatbot from scratch, teams adapt one pretrained model with prompting, retrieval, fine-tuning, or a small adapter. BERT and the GPT family are familiar examples of this shift, and the word itself came out of Stanford's 2021 report 'On the Opportunities and Risks of Foundation Models'.
Reach for one when the task needs broad language, image, audio, or coding ability and your own data is too small to train that ability from zero. The practical choice is not just which family wins a benchmark; it is whether the model's license, modalities, context limit, serving cost, and deployment options fit the product.
Gotcha: foundation does not mean factual, neutral, or ready for production. The model inherits gaps and biases from pretraining, and adaptation can trade away useful general behavior. Test the exact adapted version on a fixed eval set, and keep permissions and business rules outside the model.
Ask AI for it
Adapt google/gemma-2-2b for classifying these support tickets with Hugging Face Transformers and a PEFT LoRA adapter. Keep the base weights frozen, split the labeled data into train, validation, and test sets, report per-class precision and recall, save the adapter separately from the base model, and include one command to reproduce training plus one command to run deterministic evaluation on the held-out test set.