--- language: - en task_categories: - visual-question-answering tags: - benchmark - multimodal - embodied-ai - ground-truth configs: - config_name: default data_files: - split: test path: - metadata.jsonl - assets/** dataset_info: features: - name: image dtype: image - name: question dtype: string - name: choices list: string - name: answer dtype: string - name: answer_type dtype: string - name: auxiliary_images list: image - name: capability_name dtype: string - name: question_type_name dtype: string - name: difficulty dtype: string - name: id dtype: string - name: sample_id dtype: string - name: split dtype: string - name: capability_id dtype: string - name: question_type dtype: string - name: template_id dtype: string - name: template_variant_id dtype: string - name: option_labels list: string - name: option_texts list: string - name: options_json dtype: string - name: answer_format dtype: string - name: metric_name dtype: string - name: metric_family dtype: string - name: scorer dtype: string - name: scoring_rule dtype: string - name: vlmeval_compatible dtype: bool - name: difficulty_reason dtype: string - name: gt_fields list: string - name: metadata_json dtype: string --- # BenchForge UAV The test split contains 10,000 English visual questions over real aerial imagery. Each example links its main image and ordered auxiliary images to the full-resolution originals in this repository. - The Hugging Face ImageFolder metadata in the root links all questions to their original images in the assets directory. - The English JSONL records under data retain the same question IDs, choice labels, answers, image paths, and scoring definitions. - There are 14,866 original image paths and 16,891 image references across the test split. Stream the complete test split using Hugging Face Datasets. The browser preview can cover a subset of large ImageFolder datasets. from datasets import load_dataset dataset = load_dataset("jby20180901/BenchForge-uav", split="test", streaming=True)