VLFeedback: 80,000 Multi-Modal Instructions with GPT-4V Preference Labels

VLFeedback contains 80,000 multi-modal instructions and 320,000 model responses annotated by GPT-4V for vision-language preference learning. Developed by MMInstruction in late 2023, the dataset aggregates instructions from diverse sources to evaluate a pool of 12 different Large Vision-Language Models (LVLMs).

Use Cases

• Training reward models for RLHF using the GPT-4V preference rankings
• Benchmarking LVLM performance by comparing the four model responses provided for each instruction
• Instruction fine-tuning of vision-language models to improve alignment with AI-generated preferences

Strengths

• 80,000 multi-modal instructions
• Comparative data featuring responses from 12 distinct LVLMs
• High-quality synthetic annotations provided by GPT-4V

Limitations

• Reliance on synthetic GPT-4V annotations which may contain model-specific biases compared to human ground truth
• Potential for temporal staleness as the 12 source LVLMs are superseded by newer architectures

Provenance

Source

MMInstruction (Arxiv 2312.10665)

Collection Method

Synthetic annotation of model-generated responses using GPT-4V

Time Range

2023

Freshness

Last updated October 2024; reflects model capabilities and instructions available as of late 2023.