FineVision: Massive Open Vision-Language Data скачать в хорошем качестве

FineVision: Massive Open Vision-Language Data

In this AI Research Roundup episode, Alex discusses the paper: 'FineVision: Open Data Is All You Need' FineVision introduces a large, open, and rigorously curated vision–language corpus built to fix fragmentation, inconsistency, and contamination in public datasets for VLMs. It aggregates 200+ sources into 185 subsets with 24M samples, 17M images, 89M conversational turns, and 9.5B answer tokens, prioritizing data hygiene and safety. A semi-automated pipeline maps diverse tasks into a unified chat-style schema with LLM-assisted converters and human-in-the-loop auditing. Robust cleaning, deduplication (SSCD + cosine), and benchmark decontamination help preserve fidelity while minimizing leakage. Paper URL: https://arxiv.org/abs/2510.17269 #AI #MachineLearning #DeepLearning #VisionLanguage #VLM #OpenDataset #DataCuration #Multimodal