Visual Instruction Tuning

🤖 Plain-English Summary

Instruction tuning large language models (LLMs) using machine-generated instruction-following data has improved zero-shot capabilities on new tasks, but the idea is less explored in the multimodal field. When fine-tuned on Science QA, the synergy of LLaVA and GPT-4 achieves a new advanced accuracy of 92.53%.

🔑 Key Findings

• In this paper, we present the first attempt to use language-only GPT-4 to generate multimodal language-image instruction-following data.
• By instruction tuning on such generated data, we introduce LLaVA: Large Language and Vision Assistant, an end-to-end trained large multimodal model that connects a vision encoder and LLM for general-purpose visual and language understanding.Our early experiments show that LLaVA demonstrates impressive multimodel chat abilities, sometimes exhibiting the behaviors of multimodal GPT-4 on unseen images/instructions, and yields a 85.1% relative score compared with GPT-4 on a synthetic multimodal instruction-following dataset.
• When fine-tuned on Science QA, the synergy of LLaVA and GPT-4 achieves a new advanced accuracy of 92.53%.

🔗 Related Resources