--- dataset_info: features: - name: id dtype: string - name: index dtype: int32 - name: qtype dtype: string - name: domain dtype: string - name: subdomain dtype: string - name: question dtype: string - name: answer dtype: string - name: options dtype: string - name: option_map dtype: string - name: source_image dtype: string - name: image dtype: image task_categories: - visual-question-answering language: - en pretty_name: UNREAL Benchmark size_categories: - n<1K configs: - config_name: default data_files: - split: train path: train-*.parquet tags: - visual-reasoning - benchmark - multimodal - genesis --- # UNREAL Benchmark

Visual reasoning across unfamiliar scenes
600 questions · 5 domains · Embedded images · English

UNREAL Benchmark is an evaluation dataset for visual question answering across mathematics, natural science, perception, physics and engineering, and reasoning and logic. Each example pairs an image with a question and a reference answer. ## A look inside Six examples from the released benchmark. Click any image to view it at full resolution. The gallery uses the original image bytes, with no cropping or retouching.
Reasoning about local gravity — PHYS_0001
Physics & Engineering
Reasoning about local gravity
PHYS_0001		 Anatomy and locomotion — NATS_0001
Natural Science
Anatomy and locomotion
NATS_0001
Geometry in a visual scene — MATH_0001
Mathematics
Geometry in a visual scene
MATH_0001		 Spatial orientation and shadows — PERC_0001
Perception
Spatial orientation and shadows
PERC_0001
Recognizing anachronisms — REAS_0001
Reasoning & Logic
Recognizing anachronisms
REAS_0001		 Interpreting illumination — REAS_0004
Reasoning & Logic
Interpreting illumination
REAS_0004
## Dataset at a glance | Domain | Examples | |---|---:| | Mathematics | 259 | | Natural Science | 62 | | Perception | 91 | | Physics & Engineering | 96 | | Reasoning & Logic | 92 | | **Total** | **600** | **Question formats:** 458 open-ended (`oe`) and 142 multiple-choice (`mc`). The release contains three Parquet shards of 200 examples each. Images are embedded in the `image` column. The existing split name is `train`; the content is intended for evaluation. ## Load the benchmark ```python from datasets import load_dataset dataset = load_dataset("CauchyFan/UNREAL_Benchmark", split="train") example = dataset[0] print(example["id"], example["question"]) image = example["image"] # PIL image ``` To iterate without downloading the entire release first: ```python stream = load_dataset( "CauchyFan/UNREAL_Benchmark", split="train", streaming=True ) example = next(iter(stream)) ``` ## Fields | Field | Description | |---|---| | `id`, `index` | Sample identifier and numeric index. | | `domain`, `subdomain` | Topic labels. | | `qtype` | `oe` for open-ended or `mc` for multiple-choice. | | `question`, `answer` | Question text and reference answer. | | `options` | JSON-encoded option mapping when present; parse with `json.loads`. | | `option_map` | Human-readable labeled options when present. | | `source_image` | Original image filename. | | `image` | Embedded image decoded by Hugging Face Datasets. | For multiple-choice evaluation, present both the question and its labeled options. Keep reference answers separate from model inputs. ## Genesis model family Public checkpoints: [Genesis-Qwen3vl-4B](https://huggingface.co/CauchyFan/Genesis-Qwen3vl-4B) · [Genesis-Qwen3vl-8B](https://huggingface.co/CauchyFan/Genesis-Qwen3vl-8B) · [Genesis-Qwen3vl-32B](https://huggingface.co/CauchyFan/Genesis-Qwen3vl-32B)