README.md

metadata

license: other
license_name: mixed
license_link: LICENSE.md
language:
  - en
pretty_name: OmniTaskonomy
task_categories:
  - visual-question-answering
  - image-to-image
tags:
  - multimodal
  - visual-understanding
  - image-editing
  - omnitaskonomy
configs:
  - config_name: i2t
    default: true
    data_files:
      - split: recognition__activity_understanding__i2t
        path: data/recognition__activity_understanding__i2t/val/*.parquet
      - split: recognition__anomaly_detection__i2t
        path: data/recognition__anomaly_detection__i2t/val/*.parquet
      - split: recognition__appearance_understanding__i2t
        path: data/recognition__appearance_understanding__i2t/val/*.parquet
      - split: recognition__category_recognition__i2t
        path: data/recognition__category_recognition__i2t/val/*.parquet
      - split: recognition__ocr_and_text_recognition__i2t
        path: data/recognition__ocr_and_text_recognition__i2t/val/*.parquet
      - split: recognition__part_recognition__i2t
        path: data/recognition__part_recognition__i2t/val/*.parquet
      - split: recognition__region_semantic_recognition__i2t
        path: data/recognition__region_semantic_recognition__i2t/val/*.parquet
      - split: recognition__situation_understanding__i2t
        path: data/recognition__situation_understanding__i2t/val/*.parquet
      - split: recognition__state_recognition__i2t
        path: data/recognition__state_recognition__i2t/val/*.parquet
      - split: recognition__symbol_recognition__i2t
        path: data/recognition__symbol_recognition__i2t/val/*.parquet
      - split: recognition__visual_similarity__i2t
        path: data/recognition__visual_similarity__i2t/val/*.parquet
      - split: reconstruction__2d_spatial_relation__i2t
        path: data/reconstruction__2d_spatial_relation__i2t/val/*.parquet
      - split: reconstruction__3d_shape_reasoning__i2t
        path: data/reconstruction__3d_shape_reasoning__i2t/val/*.parquet
      - split: reconstruction__depth_understanding__i2t
        path: data/reconstruction__depth_understanding__i2t/val/*.parquet
      - split: reconstruction__lighting_understanding__i2t
        path: data/reconstruction__lighting_understanding__i2t/val/*.parquet
      - split: reconstruction__metric_3d_relation__i2t
        path: data/reconstruction__metric_3d_relation__i2t/val/*.parquet
      - split: reconstruction__multi_view_reasoning__i2t
        path: data/reconstruction__multi_view_reasoning__i2t/val/*.parquet
      - split: reconstruction__occlusion_understanding__i2t
        path: data/reconstruction__occlusion_understanding__i2t/val/*.parquet
      - split: reconstruction__orientation_understanding__i2t
        path: data/reconstruction__orientation_understanding__i2t/val/*.parquet
      - split: reorganization__2d_keypoints__i2t
        path: data/reorganization__2d_keypoints__i2t/val/*.parquet
      - split: reorganization__2d_ordering__i2t
        path: data/reorganization__2d_ordering__i2t/val/*.parquet
      - split: reorganization__connectivity__i2t
        path: data/reorganization__connectivity__i2t/val/*.parquet
      - split: reorganization__counting__i2t
        path: data/reorganization__counting__i2t/val/*.parquet
      - split: reorganization__localization__i2t
        path: data/reorganization__localization__i2t/val/*.parquet
      - split: reorganization__visual_correspondence__i2t
        path: data/reorganization__visual_correspondence__i2t/val/*.parquet
  - config_name: i2i
    data_files:
      - split: recognition__attribute_editing__i2i
        path: data/recognition__attribute_editing__i2i/train/*.parquet
      - split: recognition__object_editing__i2i
        path: data/recognition__object_editing__i2i/train/*.parquet
      - split: reconstruction__inpainting__i2i
        path: data/reconstruction__inpainting__i2i/train/*.parquet
      - split: reorganization__jigsaw__i2i
        path: data/reorganization__jigsaw__i2i/train/*.parquet
      - split: reorganization__localization__i2i
        path: data/reorganization__localization__i2i/train/*.parquet
      - split: reorganization__object_pointing__i2i
        path: data/reorganization__object_pointing__i2i/train/*.parquet
      - split: reorganization__semantic_segmentation__i2i
        path: data/reorganization__semantic_segmentation__i2i/train/*.parquet

OmniTaskonomy

OmniTaskonomy (arxiv.org/abs/2609.38079) groups visual tasks into Recognition, Reconstruction, and Reorganization. Each task and modality has its own split, named family__task__i2i or family__task__i2t. The i2t and i2i configs group the evaluation and training splits, respectively.

This release contains 9,444 I2T evaluation samples across 25 tasks and 350,000 I2I training samples across 7 tasks. I2T rows have usage="val"; I2I rows have usage="train". I2I validation data is not included. Counts and file checksums are in release_manifest.json.

Load a task

from datasets import load_dataset

data = load_dataset(
    "Wakals/OmniTaskonomy",
    name="i2t",
    split="recognition__category_recognition__i2t",
    streaming=True,
)
sample = next(iter(data))

For I2I training data, use name="i2i" and the corresponding __i2i split.

Fields

• id: unique release record; task_id, task_name, r, modality, and usage describe its task.
• input_images: an ordered image list, including for single-image tasks.
• I2T: question, options, answer, optional hint, and evaluation_prompt.
• I2I: editing_prompt and output_image.
• source_dataset, source_id, source_split, source_license, and JSON-encoded source_metadata retain provenance. repeat_index distinguishes repeated source records.

Fields that do not apply to a modality are null. I2T answers are option labels. Original I2T image bytes, image order, questions, and answers are preserved; evaluation_prompt retains the benchmark's evaluation wording, including its native CV-Bench prompt. Source annotation anomalies are retained.

I2I targets use the training reader's rendering. For damaged source files accepted by the training reader, the same decoded pixels are saved as valid PNGs; affected records are identified in source_metadata. Localization repeats the original training iterator to 50,000 records. Object pointing stores one deterministic choice from the training prompt templates; the templates remain available in source metadata.

Taskonomy source selections

For the 12 Taskonomy-derived I2I tasks, sources.json provides official download links, archive checksums, prompts, and preprocessing rules. The compressed ID lists contain 50,000 training occurrences per task; preserve their order and repeated IDs. These are source metadata files, not additional dataset splits. Taskonomy images are not included while redistribution permission is pending.

License

This dataset combines sources with different licenses. See LICENSE.md for each source's terms, including RealWorldQA's CC BY-ND 4.0 license. Each component retains its original license; no relicensing of third-party content is intended.