noun
A downstream task is a practical application that relies on the output of a pre-trained AI model, such as a language model or image generator. Unlike the initial training phase, where the model learns from vast datasets, downstream tasks apply the model to specific real-world problems, like summarizing articles, generating captions, or answering questions.
These tasks often require fine-tuning the model on a smaller, specialized dataset to improve performance. For example, a chatbot trained on general conversation data might be fine-tuned on medical knowledge to answer health-related questions. The cost and complexity depend on the task—some require minimal adjustments, while others need significant computational resources.
The main change for users is that downstream tasks make AI more accessible and practical. They allow developers to build applications without training models from scratch, saving time and money. However, the quality of the output still depends on the original model’s capabilities and the fine-tuning process. For instance, a poorly fine-tuned model might generate inaccurate medical advice, risking user trust.
A company uses a pre-trained language model to build a customer support chatbot, fine-tuning it on past service tickets to improve response accuracy.
In other languages
See also: fine-tuning · Transfer Learning