jiahaozhang2003/ncrna-pretrain-47m-cluster-split

ncRNA-pretrain-47M Cluster Split

This dataset is a train/validation/test split of syjiang/ncRNA-pretrain-47M for RNA foundation-model pretraining.

The split is deterministic and is performed at the cluster_id level, so all rows with the same cluster_id are assigned to the same split. The original columns are preserved unchanged:

id, source, rna_type, length, cluster_id, is_representative, sequence, description.

Source snapshot

• Source dataset: syjiang/ncRNA-pretrain-47M
• Source revision: 79d9b232405d0325f819de3cc8ac40983e6bbeec
• Source last modified: 2026-06-22T03:03:00.000Z
• License: CC-BY-4.0

Split rule

bucket = SplitMix64(cluster_id, seed=20260622) % 10000
0-9849    -> train
9850-9899 -> validation
9900-9999 -> test

Target fractions are 98.5% / 0.5% / 1.0% by cluster.

Split sizes

split	rows	clusters	nucleotides	parquet files	parquet bytes
train	46,442,722	20,802,298	31,994,958,841	76	8,026,715,599
validation	227,966	105,427	149,245,337	1	43,146,203
test	483,433	211,089	339,035,084	2	89,029,804

Loading

from datasets import load_dataset

ds = load_dataset("jiahaozhang2003/ncrna-pretrain-47m-cluster-split")

For pretraining, use the train split with cluster-balanced sampling if desired. Use validation for MLM checkpoint selection and K sweeps, and reserve test for final held-out evaluation/probing.