Upload 3 files

zanyi/README.md

@@ -0,0 +1,68 @@
+---
+license: other
+task_categories:
+- text-to-image
+language:
+- zh
+- en
+size_categories:
+- n<1K
+pretty_name: Knowledge-Grounded Image Synthesis Benchmark
+configs:
+- config_name: default
+  data_files:
+  - split: test
+    path: data/test-*.parquet
+---
+
+# Knowledge-Grounded Image Synthesis Benchmark
+
+200 knowledge-grounded prompts × 2 image-generation models (Gemini, GPT-Image-1) = 400 generated images.
+
+Each row of the dataset carries one prompt and the two model outputs side-by-side, so the HF dataset viewer renders **prompt + Gemini thumbnail + GPT-Image-1 thumbnail** on a single row.
+
+## Schema
+
+| column           | type        | description                                  |
+|------------------|-------------|----------------------------------------------|
+| `id`             | string      | e.g. `A-1`, `G-120`                          |
+| `category`       | string      | `A` / `B` / `C` / `D` / `G`                  |
+| `category_name`  | string      | human-readable category name                 |
+| `language`       | string      | `zh` or `en`                                 |
+| `prompt`         | string      | the input prompt                             |
+| `gemini_image`   | image       | Gemini output (jpg bytes, embedded)          |
+| `gptimage_image` | image       | GPT-Image-1 output (png bytes, embedded)     |
+
+## Categories
+
+| Code | Name                                   | Count |
+|------|----------------------------------------|-------|
+| A    | Named Entity Multi-Attribute Display   | 20    |
+| B    | Relational Binding                     | 20    |
+| C    | Ordering / Classification              | 20    |
+| D    | Event-Grounded Specific Moment         | 20    |
+| G    | Composite Knowledge Layout             | 120   |
+
+Bilingual: each category mixes Chinese (`zh`) and English (`en`) prompts.
+
+## Loading
+
+```python
+from datasets import load_dataset
+ds = load_dataset("knowledge-in-visual-synthesis/knowledge-image-synthesis", split="test")
+row = ds[0]
+row["prompt"]          # str
+row["gemini_image"]    # PIL.Image
+row["gptimage_image"]  # PIL.Image
+```
+
+## File layout
+
+```
+data/
+├── test-00000-of-00002.parquet   # rows 0..99   (~273 MB)
+└── test-00001-of-00002.parquet   # rows 100..199 (~294 MB)
+README.md
+```
+
+Image bytes are embedded directly in the parquet files; no external image directory is required.

zanyi/data/test-00000-of-00002.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:27dbbb4bd8bd797f59a4ebc86aa836cfeb8a02a0636282aa7f537e10249df2ad
+size 272636886

zanyi/data/test-00001-of-00002.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:53983228f0351401449c95a1c51f12201b02d3371e73bb9d6d16217bf805a4e0
+size 294037512