SII-GAIR-NLP/RIBOSPAN-FM-Corpus

RIBOSPAN Pretraining Corpus

The RIBOSPAN Pretraining Corpus is the RNA sequence corpus used to pretrain the RIBOSPAN model family.

The corpus combines diverse RNA sequences from RNAcentral v26.0 with quality-controlled protein-coding transcripts from Ensembl release 115 and Ensembl Genomes release 62.

After source-specific filtering, normalization, and exact deduplication, the final training corpus contains 67.6 million RNA sequences and 85.7 billion nucleotide tokens in total.

---

🧬 Data Sources

The RIBOSPAN pretraining corpus was constructed from:

• RNAcentral v26.0
• Ensembl release 115
• Ensembl Genomes release 62

RNAcentral provides broad coverage across RNA classes, while Ensembl datasets contribute quality-controlled protein-coding transcripts with complete CDS and UTR annotations across vertebrates, plants, fungi, metazoans, and protists.

RNAcentral v26.0, Ensembl release 115, and Ensembl Genomes release 62 were curated independently before merging.

---

🧩 Data Curation

Sequence Normalization

All sequences were normalized to uppercase, with U mapped to T and unsupported symbols mapped to N.

The resulting corpus therefore uses the normalized nucleotide alphabet:

A C G T N

RNAcentral

For RNAcentral, the active and inactive sequence sets were combined and exact-deduplicated by sequence using SeqKit.

Ensembl and Ensembl Genomes

Ensembl cDNA sequences from vertebrates, plants, fungi, metazoans, and protists were paired with the corresponding release-matched non-ab-initio GTF annotations and filtered to retain complete, high-quality protein-coding transcripts, followed by sequence-level exact deduplication using SeqKit.

Filtering Criteria for Ensembl Protein-Coding Transcripts

Criterion	Requirement
Annotation	Protein-coding transcript annotation and, when available, protein-coding gene annotation; neither is annotated as a pseudogene.
Completeness	CDS, start codon, stop codon, 5′ UTR, and 3′ UTR annotations are present.
Coding consistency	CDS length is positive and divisible by 3, with consistent start- and stop-codon annotations.
Sequence quality	≥ 20 nt, ≤ 5% N, and no homopolymer > 50 nt.

---

✂️ Dataset Splits

After source-specific filtering, normalization, and exact deduplication, the curated corpus was divided into training, validation, and test splits.

The validation and test sets were constructed using class-specific sampling quotas to provide broad representation across RNA classes while reducing the influence of the highly imbalanced training distribution.

For protein-coding transcripts, the sampling was further stratified by species.

The validation and test sets each contain 90,000 held-out RNA sequences.

The released dataset consists of three sequence-level splits in compressed FASTA format:

Split	File
Training	train.fasta.gz
Validation	validation.fasta.gz
Test	test.fasta.gz

The validation and test sequences are held out at the sequence level and are not included in the training split.

---

📊 Dataset Statistics

Sequence-Length and RNA-Class Distributions

The resulting sequence-length and RNA-class distributions are shown below.

Sequence-length distributions and RNA-class composition of the pretraining corpus.

RNA-Class Composition of the Pretraining Corpus and Held-Out Sets

RNA Class	Training	Validation	Test
rRNA	30,748,160	10,000	10,000
mRNA	18,316,110	10,000	10,000
tRNA	6,135,163	10,000	10,000
misc RNA	3,966,721	10,000	10,000
lncRNA	3,818,804	10,000	10,000
others	940,161	5,000	5,000
pre miRNA	804,307	5,000	5,000
snoRNA	598,283	5,000	5,000
sRNA	541,829	5,000	5,000
ncRNA	510,635	5,000	5,000
snRNA	493,584	5,000	5,000
SRP RNA	483,594	5,000	5,000
piRNA	209,734	5,000	5,000
Total	67,567,085	90,000	90,000

The training corpus follows the natural RNA-class distribution and is therefore highly imbalanced, whereas the held-out validation and test sets use class-specific sampling quotas to provide broad representation across RNA classes.

Because RNA classes differ substantially in sequence length, sequence-count proportions do not directly correspond to nucleotide proportions. Long transcript classes such as mRNA therefore contribute disproportionately to the total number of nucleotides relative to their sequence counts.

Pretraining Corpus Statistics by Split and Source

Split	Source	Sequences	Nucleotide Tokens
Training	RNAcentral v26.0	49,250,975	34,111,085,735
Training	Ensembl 115 / Ensembl Genomes 62	18,316,110	51,603,244,706
Validation	Combined	90,000	77,781,035
Test	Combined	90,000	76,893,030

The two training sources together contain 67,567,085 RNA sequences and 85,714,330,441 nucleotide tokens.

Including the validation and test splits, the released corpus contains 67,747,085 RNA sequences and 85,869,004,506 nucleotides in total.

---

🧾 Sequence Format

The corpus is distributed in compressed FASTA format.

Each FASTA record corresponds to one independent RNA sequence.

>sequence_id
ATGCTGAC...

All released sequences use the normalized nucleotide alphabet:

A C G T N

RNA uracil (U) is represented as thymine (T), consistent with the sequence normalization used during RIBOSPAN pretraining.

No [CLS], [SEP], [MASK], padding, or other model-specific tokens are stored in the released FASTA files. These tokens are introduced during model-side tokenization and training sample construction.

---

🙏 Acknowledgements

We thank RNAcentral, Ensembl, and Ensembl Genomes for providing the sequence resources used to construct the RIBOSPAN pretraining corpus.

We also acknowledge the developers of SeqKit for providing the sequence-processing toolkit used for exact sequence deduplication during corpus construction.

Users of this dataset should appropriately acknowledge and cite the original data resources in accordance with their respective terms and recommended citation practices.

---

📖 Citation

If you use the RIBOSPAN Pretraining Corpus in your research, please cite the RIBOSPAN paper:

@misc{wang2026ribospan,
  title         = {{RIBOSPAN}: A Long-Context {RNA} Foundation Model for Versatile {RNA} Modeling},
  author        = {Wang, Ziyuan and Tang, Bohao and Zhang, Fei and Han, Shuo and Liu, Pengfei},
  year          = {2026},
  eprint        = {2608.22849},
  archivePrefix = {arXiv},
  primaryClass  = {cs.LG}
}

---

RiboSpan — Single-nucleotide Pretraining Across Native-contex.