ai-forever/fusion_brain_aij2021 ? reverse-engineered prompt

Build me a single multimodal model that can handle four different tasks in one go, code translation, handwritten text recognition, zero shot object detection, and visual question answering.

I want it to feel like one shared system, not four completely separate models, so reuse as much of the core network as possible and make the shared parts actually matter for prediction. Use the ideas from this repo’s concept and baselines if they help, and if you need to look up current docs or example approaches online, do that.

Please also make the whole thing runnable for inference the way this challenge expects, so it reads the task inputs and writes four output JSON files, one for each task, in the right format. It should be able to work on the provided sample data and be easy to run from the notebook or a simple script.

If there are obvious preprocessing steps, like handling rotated handwritten images or normalizing inputs, include them too.