---
license: mit
---

# The GraphTide Real-World Dynamic Graph Datasets

GraphTide is a collection of large dynamic graph datasets in a compact binary
format for streaming graph algorithms, dynamic graph benchmarks, and temporal
graph analysis.

Download the source code from [`wheatman/GraphTide`](https://github.com/wheatman/GraphTide).
It contains the binary IO spec, Python readers/converters, source parsers, and C++ analysis code.

## Layout
```text
signal_graphs/   insert-only temporal edge streams
update_graphs/   explicit edge insert/delete streams
```

## Download
The full dataset is large, so download only the files you need when possible.
```bash
pip install -U huggingface_hub
huggingface-cli download DynamicGraphsProvider/GraphTide \
  signal_graphs/soc-youtube-growth.bin \
  --repo-type dataset --local-dir GraphTide
```
To clone everything:
```bash
git lfs install
git clone https://huggingface.co/datasets/DynamicGraphsProvider/GraphTide
```

## Binary Format
Each file starts with a 32-byte little-endian header: three `uint64` values
(`num_vertices`, `num_unique_edges`, `total_updates`), five `uint8` values
(`signal`, `vertex_id_type`, `timestamp_type`, `directed`, `weight_type`), and
three padding bytes.
Rows contain `src`, `dst`, optional `timestamp`, optional `weight`, and for
update graphs an `info` byte. The shipped files are directed, unweighted, and use
`UINT32` vertex IDs and `UINT32` timestamps.

See https://github.com/wheatman/GraphTide/blob/main/IO/README.md for details.