OFASys: Implementing Multitask Learning with Just One Line of Code

Developing generalist models through multimodal multitask learning is drawing attention, but the process of integrating and stabilizing training across diverse modalities and tasks is highly complex. There are simply too many details—data preprocessing, model construction, batch composition, and more—that engineers must resolve.

OFASys is an AI framework designed to solve this problem. Its core is a simple interface called Instruction. Users only need to write an Instruction—a template containing task-specific instructions and input information—in just one line of code.

This allows users to focus solely on task and modality design, without having to worry about complex processes like data preprocessing, model construction, and training.

Internally, the system operates by parsing the Instruction into a Task Plan. Each plan has a model hierarchy consisting of modality-specific pre/post-processors and adapters, along with a modality-agnostic Universal Model.

The Universal Model serves as a general-purpose module that fuses multimodal inputs and generates outputs, and the final output is produced as content matching the input format through adapters and post-processors.