DICS: Exploring Data Intrinsic Consistency for Visual Instruction Selection
DICS selects visual-instruction training samples by scoring how well each sample’s image, instruction, and response cohere.
The paper introduces Data Intrinsic Consistency, with separate checks for image-instruction alignment and response-instruction coherence. Its selection method balances those scores against dataset diversity under different data budgets. The authors report that DICS beats prior selection methods and can outperform full LLaVA-1.5-665K fine-tuning while using 25% of that data. They also release DICS-6M and say the method reaches 94.52% of official InternVL3-8B-Instruct performance with less than 25% of the reported training data. HF Daily Papers' note
The paper introduces Data Intrinsic Consistency, with separate checks for image-instruction alignment and response-instruction coherence. Its selection method balances those scores against dataset diversity under different data budgets. The authors report that DICS beats prior selection methods and can outperform full LLaVA-1.5-665K fine-tuning while using 25% of that data. They also release DICS-6M and say the method reaches 94.52% of official InternVL3-8B-Instruct performance with less than 25% of the reported training data. HF Daily Papers' note
score 4