---
license: apache-2.0
language:
- en
tags:
- graph
- text-attributed-graph
- node-classification
- llm
size_categories:
- 1M<n<10M
---

# GraspLLM Datasets

Pre-processed text-attributed graphs used in
[**GraspLLM**](https://github.com/Heinz217/GraspLLM):
*Towards Zero-Shot Generalization on Text-Attributed Graphs with LLMs*.

This repository hosts the **raw graph** (`processed_data.pt`) for each dataset.

## Datasets

We follow the paper's 14 benchmarks:

| Domain | Dataset | File |
|---|---|---|
| Citation | Cora, Citeseer, Pubmed, Arxiv | `cora/`, `citeseer/`, `pubmed/`, `ogbn-arxiv/` |
| E-commerce | Books-History, Ele-Computer, Ele-Photo | `history/`, `computer/`, `photo/` |
| Wikipedia | WikiCS | `wikics/` |
| Social | Reddit, Instagram | `reddit/`, `instagram/` |
| Web pages | Cornell, Texas, Washington, Wisconsin | `cornell/`, `texas/`, `washington/`, `wisconsin/` |

Each `<dataset>/processed_data.pt` is a [PyG](https://pytorch-geometric.readthedocs.io/) `Data` object containing:

- `x`: node feature tensor (placeholder; replaced by Qwen3-Embedding-8B at runtime)
- `edge_index`: `[2, E]` undirected edge list
- `y`: node labels
- `train_mask` / `val_mask` / `test_mask`
- `raw_texts`: list of raw node texts (used to compute Qwen3 features)
- `label_texts` (where applicable): the textual class names

## Usage

```bash
# Clone the GraspLLM code repo
git clone https://github.com/Heinz217/GraspLLM.git
cd GraspLLM

# Download this dataset bundle
huggingface-cli download Heinz217/GraspLLM-Datasets \
    --repo-type dataset \
    --local-dir ./dataset

# Configure dataset root (or just keep the default ./dataset)
export GRASPLLM_DATASET_ROOT=$PWD/dataset
```