Add dataset card

README.md

@@ -0,0 +1,117 @@
+---
+pretty_name: StructHub
+license: other
+license_name: structhub-mixed-sources
+language:
+- en
+task_categories:
+- image-to-text
+tags:
+- svg
+- xml
+- structured-images
+- diagrams
+- charts
+- flowcharts
+- vector-graphics
+- image-to-svg
+- multimodal
+size_categories:
+- 10K<n<100K
+---
+
+# StructHub
+
+**StructHub** is a large-scale dataset of **structured images paired with their editable
+SVG / XML source code**, built for the *image-to-SVG* (and *text-to-SVG*) task introduced
+in **Back2Struct: Making Structured Images Editable Again**. Each raster image (a diagram,
+chart, flowchart, architecture figure, UML diagram, schema, etc.) is paired with
+**object-level** vector code that explicitly encodes text, shapes, topology, and layout —
+supporting supervised and reward-based training for models that "make structured images
+editable again."
+
+- 📄 Paper: *Back2Struct: Making Structured Images Editable Again*
+- 🔗 Project page: https://pengyu965.github.io/Back2Struct.github.io/
+- 🤖 Model: [`Pengyu965/Back2Struct-Image2SVG-7B`](https://huggingface.co/Pengyu965/Back2Struct-Image2SVG-7B)
+
+## Composition
+
+| Subset | Train (full) | Train (≤ 8,192 tok) | Retention | Benchmark |
+|--------|-------------:|--------------------:|----------:|----------:|
+| **StarDiag-HQ** | 52,875 | 52,875 | 100.0% | 569 |
+| **OpenDiag**    | 27,953 | 18,800 | 67.3%  | 334 |
+| **NNArch**      | 3,293  | 1,453  | 46.6%  | 97  |
+| **Total**       | **84,121** | **73,128** | **86.9%** | **1,000** |
+
+- **StarDiag-HQ** — high-quality synthetic structured diagrams.
+- **OpenDiag** — crawled open-domain structured graphics (broad style/topic coverage).
+- **NNArch** — neural-network / system architecture diagrams (long, complex SVGs).
+
+Training uses examples with **≤ 8,192 tokens** of ground-truth SVG (73,128 pairs) for
+stable long-context generation.
+
+## StructHub Benchmark (1,000 samples)
+
+A held-out **1,000-sample** evaluation benchmark (569 StarDiag-HQ + 334 OpenDiag +
+97 NNArch), split by ground-truth SVG length into three difficulty tiers:
+
+| Tier | GT SVG length | # samples |
+|------|---------------|----------:|
+| Easy   | ≤ 2,048 tokens        | 333 |
+| Medium | 2,048 – 4,096 tokens  | 333 |
+| Hard   | > 4,096 tokens        | 334 |
+
+This enables evaluation across levels of structural complexity.
+
+## Data fields
+
+Each example contains:
+
+| Field | Type | Description |
+|-------|------|-------------|
+| `id` | string | Unique example id (source-prefixed, e.g. `starvector_…`, `crawled_…`) |
+| `image` | image | The rendered structured raster image (input) |
+| `svg` | string | Ground-truth object-level SVG / XML code (target) |
+| `source` | string | `StarDiag-HQ` \| `OpenDiag` \| `NNArch` |
+| `num_tokens` | int | Token length of the ground-truth SVG |
+| `difficulty` | string | `easy` \| `medium` \| `hard` (benchmark split only) |
+
+> Adjust the field names above to match the files you upload.
+
+## Usage
+
+```python
+from datasets import load_dataset
+
+ds = load_dataset("Pengyu965/StructHub")
+train = ds["train"]
+bench = ds["benchmark"]
+
+ex = bench[0]
+ex["image"].save("input.png")
+print(ex["svg"][:500])   # ground-truth SVG/XML
+```
+
+## Evaluation
+
+Predictions are scored with **image-level** metrics (render success rate, DINOv2 similarity,
+LPIPS, SSIM between the rendered prediction and the input image) and a **text-level**
+GPTScore over the SVG code (Structure / Flow / Style, 0–5 each; Overall 0–100). See the
+paper for the render-gated evaluation protocol.
+
+## Licensing
+
+StructHub aggregates multiple sources, including crawled open-domain graphics. Per-source
+terms may differ; the dataset is released for **research use**. Please review and respect the
+licenses of the original content. *(Confirm and update this field before publishing.)*
+
+## Citation
+
+```bibtex
+@article{yan2026back2struct,
+  title   = {Back2Struct: Making Structured Images Editable Again},
+  author  = {Yan, Pengyu and Wu, Yixin and Tian, Yunjie and Doermann, David},
+  year    = {2026},
+  note    = {Preprint}
+}
+```