semRfam

The tokenizable multimodal RNA dataset used to train yakRNA Design, a semantic multimodal RNA composer. Every record pairs an RNA sequence with three companion modalities drawn from its Rfam family, which is what lets a model be conditioned on structure, consensus, or function at generation time.

Source	Rfam 15.0 (released 2024-09)
Records	3,978,586
Families	4,028
Format	single HDF5 file, one row per sequence
Size	~2.2 GB
License	CC0 1.0 (inherited from Rfam)

Files

File	Size	Contents
semRfam.h5	~2.2 GB	the records: sequence, ss, consensus, go_terms, rfam_id, sequence_id, each a 1-D array of UTF-8 byte strings
family_lookup.json	~49 MB	Rfam family → list of row indices, used to build family-diverse batches
family_splits.json	~35 MB	the exact family-level train/val/test split used for the released model

The four modalities

For row i, the four fields are aligned position-by-position (GO terms being per-family rather than per-position):

• sequence — the RNA sequence with alignment gaps removed, in A/C/G/U/N.
• ss — the family's WUSS #=GC SS_cons consensus structure, projected onto the positions this sequence actually occupies.
• consensus — the family's #=GC RF consensus line, projected the same way. Case is meaningful (upper = highly conserved).
• go_terms — comma-separated GO terms attached to the family by Rfam's rfam2go mapping; empty for families with no GO annotation.

Because ss and consensus are projections of the family consensus onto each sequence, len(sequence) == len(ss) == len(consensus) for every row.

Splits

Splits are by family, so no Rfam family appears in more than one split:

Split	Families	Sequences
train	3,625	3,846,938
val	201	57,920
test	202	73,728

Note: the released yakRNA checkpoint was trained with data.use_all_data_for_training: true, meaning the val/test families were also included in training; the split is published so that held-out evaluation is reproducible for anyone who wants it. family_splits.json also records the split seed and the data-processing config.

Usage

git clone https://github.com/YousufAKhan/yakRNA.git && cd yakRNA
pip install -r requirements-training.txt
python data_processing/download_rfam.py --stage prebuilt   # downloads this dataset
python rna_gen/rna_gen_training.py --config configs/train_single_gpu.yaml

That places the files where the training configs expect them (training_data/processed/). See TRAINING.md for the objective, the config reference, and how to fine-tune from the released checkpoint.

Reading it directly:

import h5py
from huggingface_hub import hf_hub_download

path = hf_hub_download("MasterYster/semRfam", "semRfam.h5", repo_type="dataset")
with h5py.File(path) as f:
    print(f["rfam_id"][0].decode(), f["sequence"][0].decode())
    print(f["ss"][0].decode())
    print(f["consensus"][0].decode())
    print(f["go_terms"][0].decode())

How it was built

1. Rfam 15.0 full-region sequences, per family.
2. Each family realigned to its covariance model with Infernal 1.1.4 cmalign --outformat Pfam.
3. Alignment gaps stripped from each sequence, with SS_cons and RF sliced to the retained positions so all three stay in register.
4. GO terms joined per family from rfam2go.
5. Written to HDF5, one row per sequence.

The pipeline is in the yakRNA repository (data_processing/download_rfam.py --stage full-all followed by data_processing/rfam_data_preprocessing.py --stage all). Note that rerunning it today yields a similar but not identical dataset: Rfam full-region membership changes between releases and search runs. This dataset is the snapshot the released model was actually trained on.

Citation

@software{yakrna2026,
  author = {Khan, Yousuf},
  title  = {yakRNA Design: A semantic multimodal RNA composer},
  year   = {2026},
  url    = {https://github.com/YousufAKhan/yakRNA}
}

Rfam: Ontiveros-Palacios et al., Rfam 15: RNA families database in 2025, Nucleic Acids Research (2025). Rfam data is released under CC0 1.0.