Papers with large-scale Natural Instructions dataset
MultiInstruct: Improving Multi-Modal Zero-Shot Learning via Instruction Tuning (2023.acl-long)
Copied to clipboard
| Challenge: | Experimental results show zero-shot performance on unseen multimodal tasks . instruction tuning has yet to be explored for vision and multimodal task. |
| Approach: | They propose a multimodal instruction tuning benchmark dataset that consists of 62 diverse multimodal tasks in a unified seq-to-seq format covering 10 broad categories. |
| Outcome: | The proposed model performs well on unseen multimodal tasks and is highly scalable. |