Add dataset card and link to paper

README.md

@@ -0,0 +1,78 @@
+---
+license: mit
+task_categories:
+- image-text-to-text
+tags:
+- gui
+- benchmark
+- vision-language-models
+- multimodal
+---
+
+# FineState-Bench
+
+[FineState-Bench](https://huggingface.co/papers/2604.27974) is a comprehensive benchmark and diagnostic framework designed for fine-grained state control in GUI agents across desktop, web, and mobile environments. 
+
+This project addresses the critical evaluation gap in current GUI agent benchmarks by shifting focus from coarse-grained task completion to precise state manipulation and control capabilities.
+
+- **Paper:** [FineState-Bench: Benchmarking State-Conditioned Grounding for Fine-grained GUI State Setting](https://huggingface.co/papers/2604.27974)
+- **Repository:** [https://github.com/FengxianJi/FineState-Bench](https://github.com/FengxianJi/FineState-Bench)
+
+## Dataset Summary
+
+FineState-Bench comprises 2,209 instances (with 2,257 high-quality tasks) across three major platforms:
+- **Desktop**: Native application environments.
+- **Web**: Browser-based interactions.
+- **Mobile**: Android and mobile-style interfaces.
+
+The benchmark spans four interaction families and 23 UI component types, with each instance explicitly specifying an exact target state for fine-grained state setting. It introduces *FineState-Metrics*, a four-stage diagnostic pipeline: Localization Success Rate (SR@Loc), Interaction Success Rate (SR@Int), and Exact State Success Rates (ES-SR).
+
+## Key Features
+
+- **Fine-Grained State Control**: Focuses on precise state manipulation rather than just final task success.
+- **Diagnostic Framework**: Includes a Visual Diagnostic Assistant (VDA) to generate descriptions and localization hints for bottleneck analysis.
+- **Multi-Platform Coverage**: Comprehensive evaluation across desktop, web, and mobile environments.
+
+## Usage
+
+You can use the benchmark by cloning the official repository and following the evaluation scripts. Below are examples of how to run the evaluation:
+
+### Desktop Evaluation
+```bash
+# Run desktop evaluation
+./desktop.sh
+
+# Or use Python directly
+python -m evaluation.benchmark --platform desktop --model GUI-R1-7B --limit 10
+```
+
+### Web Evaluation
+```bash
+# Run web evaluation
+./web.sh
+
+# Or use Python directly  
+python web.py --model Holo1-7B --limit 5 --scenario 1
+```
+
+### Mobile Evaluation
+```bash
+# Run mobile evaluation
+./mobile.sh
+
+# Or use Python directly
+python mobile.py --model SpiritSight-Agent-8B --limit 10
+```
+
+## Citation
+
+If you use FineState-Bench in your research, please cite the following:
+
+```bibtex
+@article{ji2024finestate,
+  title={FineState-Bench: Benchmarking State-Conditioned Grounding for Fine-grained GUI State Setting},
+  author={Ji, Fengxian and others},
+  journal={arXiv preprint arXiv:2604.27974},
+  year={2024}
+}
+```