Evian: Towards Explainable Visual Instruction-tuning Data Auditing

📰 ArXiv cs.AI

Learn how Evian improves data auditing for Large Vision-Language Models by providing explainable visual instruction-tuning, enabling more accurate model training

advanced Published 23 Apr 2026

Action Steps

Apply Evian to existing datasets to identify nuanced semantic flaws
Configure data filtering methods to prioritize visual fidelity and instruction-following capability
Test Evian's explainable visual instruction-tuning on a subset of data to evaluate its effectiveness
Compare Evian's performance to traditional data filtering methods
Use Evian to refine training data and improve Large Vision-Language Model accuracy

Who Needs to Know This

Data scientists and machine learning engineers working with Large Vision-Language Models can benefit from Evian's explainable data auditing to improve model performance and reduce errors

Key Insight

💡 Explainable data auditing is crucial for identifying and addressing nuanced semantic flaws in Large Vision-Language Model training data

Key Takeaways

Learn how Evian improves data auditing for Large Vision-Language Models by providing explainable visual instruction-tuning, enabling more accurate model training

Full Article

Title: Evian: Towards Explainable Visual Instruction-tuning Data Auditing

Abstract:
arXiv:2604.20544v1 Announce Type: cross Abstract: The efficacy of Large Vision-Language Models (LVLMs) is critically dependent on the quality of their training data, requiring a precise balance between visual fidelity and instruction-following capability. Existing datasets, however, are plagued by inconsistent quality, and current data filtering methods rely on coarse-grained scores that lack the granularity to identify nuanced semantic flaws like logical fallacies or factual errors. This create

Read full paper → ← Back to Reads