---
title: Puffin
canonical_url: "https://www.modelscope.cn/datasets/User040218/Puffin"
md_url: "https://www.modelscope.cn/datasets/User040218/Puffin.md"
repository: User040218/Puffin
last_updated: 2026-11-01
license: other
storage_size: "49 GB"
downloads: 18370
stars: 0
---

# Puffin

> Puffin - User040218 在 ModelScope 开源的数据集。Puffin human/mouse 4,650 bp sequences with transcription initiation targets; the web preview shows a row index.

User040218/Puffin 是 ModelScope 魔搭社区上的数据集，存储大小 49 GB，采用 other 许可。

- **Repository**: User040218/Puffin
- **License**: other
- **Storage size**: 49 GB
- **Downloads**: 18370
- **Stars**: 0
- **Last updated**: 2026-11-01

Source: https://www.modelscope.cn/datasets/User040218/Puffin

---

# Puffin

**Human and mouse 4,650 bp sequences with transcription initiation signal targets** — regression.

Puffin studies the sequence basis of transcription initiation in the human genome and predicts multiple initiation signals at base resolution. The authors release the resource files needed to reproduce the manuscript analyses on Zenodo.

This release contains the human and mouse training data: each example is a 4,650 bp sequence with 46,500 targets for human and 9,300 for mouse, using the authors' partitions (human train/dev/heldout, mouse train/dev); 80,000 rows. The previewable subsets `index_human` and `index_mouse` are row indices; use the code below to load the native files.

Links: [Puffin manuscript code (GitHub)](https://github.com/jzhoulab/puffin_manuscript), [Resource files (Zenodo 7954971)](https://zenodo.org/records/7954971).

## Subsets (2)

| Subset | Splits (rows) | Description |
|---|---|---|
| `index_human` | dev 1,591 · heldout 2,999 · train 35,410 | Human row index |
| `index_mouse` | dev 4,053 · train 35,947 | Mouse row index |

80,000 rows in total.

## Format

Parquet. Main columns:

| Column | Type | Description |
|---|---|---|
| `species` / `chromosome` / `tss` / `strand` | string / int64 | Species and TSS coordinates (index) |
| `start` / `end` | int64 | Sequence window coordinates (index) |
| `source_file` / `source_row` | string / int64 | Native file and row (index) |

## Usage

Example with [DNALLM](https://github.com/zhangtaolab/DNALLM): subset `human`, regression, `num_labels=46500`.

```python
import json
from modelscope import snapshot_download
from dnallm import DNADataset

repo = "User040218/Puffin"
root = snapshot_download(repo_id=repo, repo_type="dataset", allow_patterns=["source/native-file-map.json"])
files = json.load(open(f"{root}/source/native-file-map.json", encoding="utf-8"))["human"]
root = snapshot_download(repo_id=repo, repo_type="dataset",
                         allow_patterns=[f["path"] for split in files.values() for f in split])
data = DNADataset.load_local_data(
    {split: [f"{root}/{f['path']}" for f in items] for split, items in files.items()},
    seq_col="sequence", label_col="targets", max_length=4652)
```

The index subsets shown in the web preview are for browsing and locating rows only; the code above downloads and reads the native files listed in `source/native-file-map.json`.

For mouse, use `mouse` (9,300 targets).

## Notes

- Human and mouse targets have different dimensions and cannot be trained together.

## Source

If you use this dataset, please cite:

- Dudnyk K, Cai D, Shi C, Xu J, Zhou J. Sequence basis of transcription initiation in the human genome. Science 2024;384(6694):eadj0116. doi:10.1126/science.adj0116

## License

Declared as `other`: use of the data follows the terms of the original data providers listed in Source.