Update README.md

README.md

@@ -53,10 +53,19 @@ data/
 
 ## Data Fields
 
-Each JSON file contains structured evaluation data including:
-- **question**: The evaluation question or prompt
-- **answer**: Ground truth answer
-- **metadata**: Scene, subtask, and category information
+Each JSON sample (one file per scene) contains the following core fields:
+
+- **`task`** / **`scene`** / **`room`**: Task type, scene name, and room name
+- **`qa`**: Primary QA object containing `question` (question text), `answer_A/C` (answers), `choices_A/C` (answer options), and `view_file` (corresponding image filename)
+- **`qa_per_view`**: A list of QA pairs expanded per viewpoint — one entry per rendered camera angle
+- **`objects`**: List of objects in the scene, each with `cat` (category), `model` (model ID), and `scale` (scale factor)
+- **`stage0_initial_views`**: List of rendered viewpoints, each containing camera pose and the corresponding image filename (`filename`)
+- **`hierarchy`**: Describes the spatial relationships between objects (e.g., stacking, containment)
+- **`initial_poses`**: Initial 3D pose of each object — `position` (xyz) and `quaternion_xyzw`
+- **`orbital_geometry`**: Parameters of the orbital camera setup (center, radius, height, number of views)
+- **`seed`** / **`run_idx`**: Random seed and run index for reproducibility
+
+> Note: Each JSON file contains the full field schema shared across all 10 task categories. Fields irrelevant to the current task are set to `null` — this is expected behavior.
 
 ## Usage