Featured AI Update 🛡️🛡️🛡️
NVIDIA AI introduces Eagle 2, a VLM designed with a structured, transparent approach to data curation and model training. Eagle 2 offers a fresh approach by prioritizing openness in its data strategy. Unlike most models that only provide trained weights, Eagle 2 details its data collection, filtering, augmentation, and selection processes. This initiative aims to equip the open-source community with the tools to develop competitive VLMs without relying on proprietary datasets.
Eagle2-9B, the most advanced model in the Eagle 2 series, performs on par with models several times its size, such as those with 70B parameters. By refining post-training data strategies, Eagle 2 optimizes performance without requiring excessive computational resources.
Eagle2-9B achieves 92.6% accuracy on DocVQA, surpassing InternVL2-8B (91.6%) and GPT-4V (88.4%). In OCRBench, Eagle 2 scores 868, outperforming Qwen2-VL-7B (845) and MiniCPM-V-2.6 (852), showcasing its text recognition strengths. MathVista performance improves by 10+ points compared to its baseline, reinforcing the effectiveness of the three-stage training approach. ChartQA, OCR QA, and multimodal reasoning tasks show notable improvements, outperforming GPT-4V in key areas…..
