jiahaozhang2003/beacon-spliceai

BEACON — SpliceAI

Sequence-based splice prediction data; labels are preserved as strings to retain leading zeros.

Official data from the shared BEACON/RNABenchmark Drive folder: https://drive.google.com/drive/folders/19ddrwI8ycvIxkgSV3gDo_VunLofYd4-6?hl=en.

This repository is the standardized Hugging Face publication of the official task data. The data/ directory is the canonical viewer-friendly layer, and the original file contents and source names are preserved for reproducibility. For SSP, .npy targets are deterministically sharded into subdirectories to comply with the Hugging Face 10,000-files-per-directory limit; the original unsharded layout remains available in the GPUHub raw staging area. beacon_manifest.json records source file IDs, checksums, sizes, split counts, schema, and processing notes. Interrupted-upload artifacts ending in .partial are retained only in the GPUHub raw staging area and excluded from the published repository.

Canonical splits

Split	Rows	File
train	144628	data/train.parquet
validation	18078	data/validation.parquet
test	16505	data/test.parquet

Standardized schema

Field	Arrow type
sequence	large_string
labels	large_string

Standardization notes

• Canonical split names are train, validation, and test; source val.csv/val_1.json is published as validation.
• The official raw directory tree is preserved alongside the standardized data/*.parquet files.
• Source file names and target paths are retained in beacon_manifest.json for provenance.
• The labels column is stored as string to preserve leading zeros in binary label strings.

Usage

from datasets import load_dataset

dataset = load_dataset("jiahaozhang2003/beacon-spliceai")

Reproducibility

This repository was generated on GPUHub from the public Drive source with the RiboLoop downstream benchmark data-staging script. No model outputs or training artifacts are included.