DICS: Exploring Data Intrinsic Consistency for Visual Instruction Selection

Open paper
Benchmark Model Rank Results
object-hallucination-evaluation-on-popeDICS (LLaVA-1.5-7B, 25% LLaVA-665K)#28F1: 84.91
multimodal-reasoning-on-mmmuDICS (Qwen2-VL-7B, LLaVA-1.5-7B selector)#91Accuracy: 52

Rank counts only results with a code link.