---
title: RELAY
canonical_url: "https://www.modelscope.cn/datasets/YuzeSun/RELAY"
md_url: "https://www.modelscope.cn/datasets/YuzeSun/RELAY.md"
repository: YuzeSun/RELAY
last_updated: 2026-10-22
license: other
storage_size: "239 GB"
downloads: 231
stars: 0
---

# RELAY

> RELAY - YuzeSun 在 ModelScope 开源的数据集。Raw single-cell datasets for RELAY, organized under data/<dataset>/raw/. Original dataset licenses apply.

YuzeSun/RELAY 是 ModelScope 魔搭社区上的数据集，存储大小 239 GB，采用 other 许可。

- **Repository**: YuzeSun/RELAY
- **License**: other
- **Storage size**: 239 GB
- **Downloads**: 231
- **Stars**: 0
- **Last updated**: 2026-10-22

Source: https://www.modelscope.cn/datasets/YuzeSun/RELAY

---

# RELAY prepared datasets

Prepared inputs for **RELAY: Relational In-Context Learning for Cellular Perturbation Prediction**.

Each `data/<dataset>/` directory contains five files: `condition_counts.csv`, `cell_index.npz`, `cell_index.json`, `lognorm_hvg2000.float16.npy`, and `lognorm_hvg2000.json`. The condition table records train/context/test/control assignments; `cell_index.json` stores the label mappings, and `lognorm_hvg2000.json` stores the ordered 2,000-gene list.

| Dataset directory | Published source split | Expression matrix shape |
| --- | --- | --- |
| `PBMC_2D` | PBMC 2D, split 1 | 9,697,974 × 2,000 |
| `PBMC_3D` | PBMC 3D, split 3 | 9,697,974 × 2,000 |
| `Replogle_Nadig` | Replogle–Nadig 2D, split 1 | 966,728 × 2,000 |
| `Tahoe_100M` | Tahoe-100M 2D, split 2 | 43,624,985 × 2,000 |

Run from the RELAY code repository root:

```bash
pip install modelscope-hub
ms-hub download YuzeSun/RELAY --repo-type dataset --local-dir . --include "data/PBMC_2D/*"
```