Improve dataset card: Add metadata, paper and code links

This PR enhances the dataset card for the DemoDiff Downstream Context Data.

It adds:
- `task_categories: ['graph-ml']` to the metadata.
- Relevant `tags` such as `chemistry`, `molecular-design`, `in-context-learning`, and `drug-discovery` to improve discoverability.
- `language: ['en']` metadata for clarity.
- A direct link to the associated paper ([Graph Diffusion Transformers are In-Context Molecular Designers](https://huggingface.co/papers/2510.08744)).
- A link to the main Github repository ([https://github.com/liugangcode/DemoDiff](https://github.com/liugangcode/DemoDiff)).

These changes provide more context and improve the discoverability and utility of the dataset on the Hugging Face Hub.

README.md

@@ -1,6 +1,22 @@
+---
+task_categories:
+  - graph-ml
+tags:
+  - chemistry
+  - molecular-design
+  - in-context-learning
+  - drug-discovery
+language:
+  - en
+---
+
+---
+Paper: [Graph Diffusion Transformers are In-Context Molecular Designers](https://huggingface.co/papers/2510.08744)
+Code: https://github.com/liugangcode/DemoDiff
+---
 # DemoDiff Downstream Context Data
 
-This dataset contains context data for the DemoDiff project, organized by task names.
+This dataset contains context data for the DemoDiff project, a diffusion-based molecular foundation model for in-context inverse molecular design. It provides contextual examples to guide molecular generation, enabling few-shot molecular design across diverse chemical tasks without task-specific fine-tuning.
 
 ## Structure
 
@@ -17,4 +33,4 @@ tasks/
 │   ├── medium.csv
 │   └── negative.csv
 └── ... (other tasks)
-```
+```