Improve dataset card: Add task category, paper, code, project links, and usage

This PR significantly improves the dataset card for SGP-GenBench. It adds the `text-to-image` task category, links to the paper, GitHub repository, and project page.

It also includes a detailed introduction, information about the dataset structure, a sample usage section demonstrating how to download the data and evaluate a model, and the academic citation.

README.md

@@ -1,3 +1,91 @@
----
-license: cc-by-nc-4.0
----
+---
+license: cc-by-nc-4.0
+task_categories:
+- text-to-image
+tags:
+- graphics-programming
+- svg
+- benchmark
+- llm-evaluation
+---
+
+# SGP-GenBench: Symbolic Graphics Programming Benchmark
+
+This repository contains the dataset for **SGP-GenBench**, a comprehensive benchmark introduced in the paper [Symbolic Graphics Programming with Large Language Models](https://huggingface.co/papers/2509.05208).
+
+*   **Paper**: [https://huggingface.co/papers/2509.05208](https://huggingface.co/papers/2509.05208)
+*   **Code**: [https://github.com/Sphere-AI-Lab/SGP-RL](https://github.com/Sphere-AI-Lab/SGP-RL)
+*   **Project Page**: [https://spherelab.ai/SGP-Gen/](https://spherelab.ai/SGP-Gen/)
+
+## Introduction
+
+Large language models (LLMs) excel at program synthesis, yet their ability to produce symbolic graphics programs (SGPs) that render into precise visual content remains underexplored. This work studies symbolic graphics programming, where the goal is to generate an SGP from a natural-language description. This task also serves as a lens into how LLMs understand the visual world by prompting them to generate images rendered from SGPs. Among various SGPs, our paper sticks to scalable vector graphics (SVGs).
+
+SGP-GenBench is designed to evaluate LLMs on symbolic graphics programming, covering object fidelity, scene fidelity, and compositionality (attribute binding, spatial relations, numeracy). The benchmark reveals notable shortcomings in current models and serves as a tool for improving LLMs' ability to generate SGPs, particularly through reinforcement learning with verifiable rewards.
+
+## Dataset Structure
+
+SGP-GenBench consists of several datasets for training and evaluation:
+
+*   **Training Datasets**:
+    *   **COCO 2017**: Used for images and captions.
+    *   **`svg-gen-70k.jsonl`**: A specialized SVG training dataset.
+*   **Evaluation Datasets**:
+    *   **SGP-Object**: For evaluating object-level performance.
+    *   **SGP-CompBench**: For evaluating compositional capabilities.
+
+## Sample Usage
+
+To get started with the SGP-GenBench dataset, follow these steps:
+
+### 1. Installation
+
+First, set up the environment and install necessary packages as described in the GitHub repository:
+
+```bash
+conda env create -n sgp_gen -f environment.yml
+conda activate sgp_gen
+pip install vllm==0.7.2 && pip install oat-llm==0.0.9
+git clone [email protected]:Sphere-AI-Lab/SGP-RL.git
+cd SGP-RL
+pip install -e .
+pip install cairosvg openai-clip lxml
+```
+
+### 2. Prepare Datasets
+
+All required training and evaluation datasets can be automatically downloaded using the provided script:
+
+```bash
+bash prepare_data.sh
+```
+
+Alternatively, you can manually download `svg-gen-70k.jsonl` and `SGP-Object.json` from the Hugging Face Hub links mentioned in the GitHub README or this dataset card's files section.
+
+### 3. Evaluation on SGP-GenBench
+
+After preparing the data and potentially training a model (see the GitHub repository for RL training instructions), you can evaluate a model's performance on SGP-GenBench:
+
+```bash
+# Modify YOUR_MODEL_PATH to point to your trained model checkpoint
+python evaluate_svg_model.py --model_path YOUR_MODEL_PATH
+```
+
+The sampled responses will be stored in `./evaluation_results`. To print out important metrics like DINO-score, CLIP-score, and Diversity, use:
+
+```bash
+bash print_results.sh
+```
+
+## Citation
+
+If you use this dataset in your research, please cite the following paper:
+
+```bib
+@article{chen2025symbolic,
+  title={Symbolic Graphics Programming with Large Language Models}, 
+  author={Yamei Chen and Haoquan Zhang and Yangyi Huang and Zeju Qiu and Kaipeng Zhang and Yandong Wen and Weiyang Liu},
+  journal={arXiv preprint arXiv:2509.05208},
+  year={2025}
+}
+```