---
license: apache-2.0
language:
- en
pretty_name: OmniTaskonomy Recipe Data
task_categories:
- image-to-image
- visual-question-answering
tags:
- omnitaskonomy
- multimodal
- visual-generation
- visual-understanding
configs:
- config_name: jigsaw
  default: true
  data_files:
  - split: train
    path: data/jigsaw/train/*.parquet
  - split: val
    path: data/jigsaw/val/*.parquet
- config_name: zoomin
  data_files:
  - split: train
    path: data/zoomin/train/*.parquet
  - split: val
    path: data/zoomin/val/*.parquet
- config_name: video_unshuffle
  data_files:
  - split: train
    path: data/video_unshuffle/train/*.parquet
  - split: val
    path: data/video_unshuffle/val/*.parquet
- config_name: rotate_qa
  data_files:
  - split: train
    path: data/rotate_qa/train/*.parquet
  - split: val
    path: data/rotate_qa/val/*.parquet
- config_name: counting
  data_files:
  - split: train
    path: data/counting/train/*.parquet
  - split: val
    path: data/counting/val/*.parquet
- config_name: visgym_colorization
  data_files:
  - split: train
    path: data/visgym_colorization/train/*.parquet
  - split: val
    path: data/visgym_colorization/val/*.parquet
---

# OmniTaskonomy (arxiv.org/abs/2609.38079) Recipe Data

Paired image-to-image (I2I) and image-to-text (I2T) tasks for the R1–R6 training recipes and gradient analysis in [OmniTaskonomy](https://omni-taskonomy.github.io/).
Each of the six subsets has `train` and `val` splits. One row contains both objectives for the same task instance.

```python
from datasets import load_dataset

data = load_dataset("Wakals/OmniTaskonomy_Recipe_Data", "jigsaw", split="train", streaming=True)
sample = next(iter(data))
```

## Fields

- `id`, `task`, `split`: unique record ID, subset name, and `train` or `val`.
- `i2i_input_image`, `i2i_prompt`, `i2i_output_image`: image-generation supervision.
- `i2t_input_image`, `i2t_prompt`, `i2t_answer`: understanding supervision. Its input can differ from the I2I input.
- `source_id`, `source_dataset`, `metadata`: source identity and JSON-encoded task parameters, original split, and provenance.

Images are embedded in the Parquet files. The prompt summaries below are abbreviated; each row stores the full prompts and native answer text.

## Jigsaw

`jigsaw`: **100,000 train / 1,000 val**.
Input: a shuffled 2×2 image puzzle. I2I prompt: “Rearrange the shuffled image patches to reconstruct the original image.” Output: the reconstructed image.
I2T prompt asks for the patch order; output: `('reorder', [i0, i1, i2, i3])`, using indices **0–3** from top-left to bottom-right. I2I images use the training loader's square edge padding.

## Zoomin

`zoomin`: **100,000 train / 1,000 val**.
Input: an original image and four shuffled zoomed views. I2I prompt: “Rearrange the zoomed-in views so they are ordered from least to most zoomed.” Output: the reordered view strip.
I2T prompt asks for the same ordering; output: `('reorder', [i0, i1, i2, i3])`, using view indices **1–4**.

## Video Unshuffle

`video_unshuffle`: **30,000 train / 1,000 val**.
Input: four shuffled frames from a synthetic 3D video. I2I prompt asks to restore chronological order; output: the reordered frame strip.
I2T prompt includes the action description and asks for the frame order; output: `('reorder', [i0, i1, i2, i3])`, using frame indices **1–4**. Validation contains only the original **OOD** set.

## Rotate QA

`rotate_qa`: **50,000 train / 7,000 val**.
I2I input: the canonical COCO image, center-cropped and resized to 512×512. Prompt: “Rotate the image {degrees} degrees clockwise.” Output: the rotated image.
I2T input: that rotated image. Prompt: the stored multiple-choice rotation question; output: its original answer text. `metadata.correct_choice` stores the answer letter. Validation uses 1,000 source images with seven rotations each.

## Counting

`counting`: **50,000 train / 1,000 val**.
Input: an image and a target category. I2I prompt: “Mark all {category} in this image by placing a dot on each instance.” Output: the image with instance markers.
I2T prompt: “How many {category} are in this image? Answer with a number.” Output: the integer count. The native prompt variants and source category are retained in `metadata`.

## VisGym Colorization

`visgym_colorization`: **30,000 train / 833 val**, from `colorization_new`.
I2I input: an image with a circular region masked in gray. Prompt asks to inpaint that region while preserving the surrounding image and black outline; output: the restored image.
I2T input includes four color-wheel options beside the masked image. Prompt asks which option matches the hidden hue; output: `A`, `B`, `C`, or `D`.

## Sources and license

Jigsaw, Zoomin, and Counting use [VisGym](https://huggingface.co/datasets/VisGym/visgym_data) episodes; Colorization uses the current paired VisGym-derived pool. Rotate QA uses [COCO](https://cocodataset.org/); Video Unshuffle contains project-generated 3D scenes.

The repository retains its Apache-2.0 declaration. This does not relicense third-party images: their original source terms continue to apply, including [COCO image terms](https://cocodataset.org/#termsofuse) and the underlying VisGym image sources.
Original sample order and split membership are preserved, including existing overlaps. Counting's 1,000 validation examples also occur in its training pool. Counts and file checksums are listed in [release_manifest.json](release_manifest.json).