Abstract
We study the problem of retrieval with instructions, where users of aretrieval system explicitly describe their intent along with their queries,making the system task-aware. We aim to develop a general-purpose task-awareretrieval systems using multi-task instruction tuning that can followhuman-written instructions to find the best documents for a given query. Tothis end, we introduce the first large-scale collection of approximately 40retrieval datasets with instructions, and present TART, a multi-task retrievalsystem trained on the diverse retrieval tasks with instructions. TART showsstrong capabilities to adapt to a new task via instructions and advances thestate of the art on two zero-shot retrieval benchmarks, BEIR and LOTTE,outperforming models up to three times larger. We further introduce a newevaluation setup to better reflect real-world scenarios, pooling diversedocuments and tasks. In this setup, TART significantly outperforms competitivebaselines, further demonstrating the effectiveness of guiding retrieval withinstructions.