amazon-agi/Amazon-Nova-1.0-Canvas-evals

Amazon Nova Canvas Evaluation数据集用于评估Nova Canvas模型，使用了贪婪解码方法，并设置了最大生成长度为1600个tokens。Nova Canvas Human Evaluation Prompts是一个包含约1000个提示的A/B测试提示集，这些提示来自多个数据集，涵盖了多种类别，如人类、风景、自然场景、室内环境、创意主题和艺术主题等。

This dataset is used to evaluate the Nova Canvas model, containing approximately 1,000 prompts designed to capture customer usage of text-to-image models. The prompt set is sourced from datasets such as MSCOCO, Drawbench, OpenParti, DALL.E 3 Eval, and DOCCI, covering a wide range of categories including humans, landscapes, natural scenarios, indoor environments, creative themes, artistic themes, and more. A few prompts were randomly selected and repeated to obtain additional data points on the quality of the model.