Add dataset card

README.md

@@ -1,33 +1,64 @@
 ---
-dataset_info:
-  features:
-  - name: filename
-    dtype: string
-  - name: svg
-    dtype: string
-  - name: text
-    dtype: string
-  - name: dataset
-    dtype: string
-  splits:
-  - name: train
-    num_bytes: 532406428
-    num_examples: 85170
-  - name: test
-    num_bytes: 65293971
-    num_examples: 10599
-  - name: val
-    num_bytes: 66266242
-    num_examples: 10755
-  download_size: 224374372
-  dataset_size: 663966641
-configs:
-- config_name: default
-  data_files:
-  - split: train
-    path: data/train-*
-  - split: test
-    path: data/test-*
-  - split: val
-    path: data/val-*
+annotations_creators:
+- machine-generated
+- found
+language:
+- en
+license: other
+pretty_name: TobiaSVG
+size_categories:
+- 100K<n<1M
+source_datasets:
+- extended|QijiaHe/VFIG-Data
+- extended|yoavf/svg-animal-illustrations
+- extended|xingxm/SVGX-Core-250k
+task_categories:
+- text-generation
+tags:
+- svg
+- vector-graphics
+- text-to-svg
+- image-to-svg
 ---
+
+# TobiaSVG
+
+TobiaSVG contains 106,524 text-SVG pairs for text-to-SVG generation and
+image-to-SVG vectorization. Raster images are not stored; SVGs can be rendered
+when examples are loaded.
+
+## Sources And Splits
+
+| Subset | Source | Rows | License |
+| --- | --- | ---: | --- |
+| `vfig_diagrams` | [VFIG-Data](https://huggingface.co/datasets/QijiaHe/VFIG-Data) | 59,276 | ODC-BY 1.0 |
+| `vfig_shapes` | [VFIG-Data](https://huggingface.co/datasets/QijiaHe/VFIG-Data) | 5,832 | ODC-BY 1.0 |
+| `svgx_core` | [SVGX-Core-250k](https://huggingface.co/datasets/xingxm/SVGX-Core-250k) | 40,000 | CC BY-NC 4.0 |
+| `animal_illustrations` | [SVG Animal Illustrations](https://huggingface.co/datasets/yoavf/svg-animal-illustrations) | 1,416 | CC0 1.0 |
+
+Splits contain 85,170 training, 10,599 test, and 10,755 validation rows.
+Identical clean SVGs are kept in the same split.
+
+## Fields
+
+- `filename`: original filename.
+- `svg`: clean SVG source.
+- `text`: text description or prompt.
+- `dataset`: source subset listed above.
+
+## Processing
+
+SVGs were normalized, parsed, and rendered; invalid rows were removed. VFIG
+rows additionally required at least 40% structural primitives and at most 50
+complex shapes. Existing animal prompts and SVGX Qwen captions were retained.
+VFIG captions were generated from rendered images with
+`gemini-3.1-flash-lite`. Rows were sampled with seed 42 and split by clean SVG
+identity to prevent target leakage.
+
+## License
+
+This is a mixed-license dataset. The `dataset` field maps each row to the source
+table above. SVGX rows are CC BY-NC 4.0, so the complete dataset is for
+noncommercial use. Animal illustrations are CC0 1.0. VFIG is ODC-BY 1.0, which
+licenses the database but may not cover rights in every individual item; review
+the [ODC-BY terms](https://opendatacommons.org/licenses/by/1-0/) and source card.