---
configs:
- config_name: byx
  data_files:
  - split: v1
    path: byx/v1-*
  - split: v2
    path: byx/v2-*
- config_name: yuner
  data_files:
  - split: v1
    path: yuner/v1-*
  - split: v2
    path: yuner/v2-*
- config_name: zanyi
  data_files:
  - split: v1
    path: zanyi/v1-*
  - split: v2
    path: zanyi/v2-*
  - split: v3
    path: zanyi/v3-*
- config_name: jiayu
  data_files:
  - split: v1
    path: jiayu/v1-*
  - split: v2
    path: jiayu/v2-*
  - split: v3
    path: jiayu/v3-*
- config_name: sherry
  data_files:
  - split: v1
    path: sherry/v1-*
  - split: v2
    path: sherry/v2-*
- config_name: yujunz
  data_files:
  - split: v1
    path: yujunz/v1-*
---

# Knowledge in Visual Synthesis

This dataset contains prompt–image examples for evaluating and studying
knowledge-intensive visual synthesis. Samples are organized by contributor as
dataset subsets (configs), with each upload version exposed as a split.

## Dataset structure

| Subset | Splits |
| --- | --- |
| `byx` | `v1`, `v2` |
| `yuner` | `v1`, `v2` |
| `zanyi` | `v1`, `v2`, `v3` |
| `jiayu` | `v1`, `v2`, `v3` |
| `sherry` | `v1`, `v2` |
| `yujunz` | `v1` |

The `byx/v1` split contains 140 unique prompts and 300 generated images. For
prompts with multiple generations, images are assigned by occurrence order:
the first image is stored in `nano_banana`, the second in `gpt_image2`, and the
third in `seed_dream4d5`. The three image columns contain 140, 130, and 30
images respectively. Each row also includes structured sample metadata.

The `byx/v2` split is a curated Nano Banana-only version of `byx/v1` with 46
selected Viewer rows removed. It contains 94 unique prompts and 94 images in
the `nano_banana` column; the `gpt_image2` and `seed_dream4d5` columns are not
included in this split.


The `yujunz/v1` split contains 21 prompts spanning knowledge-graphic tasks (periodic-table grids, planet-density charts, solar-term clock faces, sudoku/score reasoning probes, and more), each with a `gpt_image2` reference image and structured metadata (`category`, `difficulty`, `checklist`, and per-image `eval` scores).

## Usage

```python
from datasets import load_dataset

dataset = load_dataset(
    "knowledge-in-visual-synthesis/v1",
    "byx",
    split="v1",
)
```

Choose another subset and split from the table above to load a different
contributor version.