DICS Metric for Visual Instruction Selection Consistency
August 30, 2026
DICS introduces Data Intrinsic Consistency, a self-scoring metric that uses Visual Information Consistency (VIC) and Response Information Consistency (RIC) to select optimal VLM training subsets. The method prioritizes samples with high internal alignment between images, instructions, and responses rather than relying solely on distribution diversity.
HOW THIS AFFECTS YOU
●
builderThis provides a method for more effective fine-tuning of vision-language models on large datasets.
●
researcherYou can improve VLM training efficiency by filtering for higher-quality, internally consistent data samples.