| # CLEVR-Hard | |
| Rendered with `generation/super_clevr/` (generic CLEVR primitives — | |
| cubes, spheres, cylinders). Each question has two variants: | |
| - **original**: base scene rendering | |
| - **distractors**: non-protected attributes randomized | |
| Generation metadata: | |
| ```json | |
| { | |
| "seed": 42, | |
| "num_val_scenes": 50, | |
| "num_test_scenes": 0, | |
| "num_objects": 6, | |
| "questions_per_scene": 10, | |
| "active_attrs": [ | |
| "color", | |
| "shape", | |
| "material", | |
| "size", | |
| "texture" | |
| ], | |
| "variants": [ | |
| "original", | |
| "distractors" | |
| ], | |
| "image_width": 480, | |
| "image_height": 320, | |
| "render_num_samples": 64 | |
| } | |
| ``` | |
| ## Fields | |
| - `id`, `scene_id`, `variant` (`original` / `distractors`) | |
| - `image` (PIL), `question`, `answer`, `question_type` | |
| - `protected`, `forbidden`, `identifying_filters` (JSON strings) | |