FILES.md

File guide

All data files are under data/. They are byte-identical to the source assets used by the BEND 2.0 evaluation package at revision bca34753d2993f6f23c7de25cc41fb8523656bbd.

The files have different schemas. Download them individually with hf_hub_download, or together with snapshot_download. The dataset viewer provides separate previews of the five Parquet tables; its data split is only a display grouping. Use each task's stored assignments and evaluation rules.

Source archives

File	Size	Contents
opensplice.tsv.gz	425.5 MB	Unchanged OpenSplice version-5 experimental table. BEND prepares 590,104 variants in 466 genes, retaining reference and variant sequences, measured effects, uncertainties and gene identifiers.
siraj.zip	93.3 MB	Unchanged Siraj saturation-mutagenesis archive. BEND prepares 340,584 mutation/cell records in 128 elements; pairing HepG2 and K562 gives 170,292 cell contrasts.
hg38.2bit	835.4 MB	Unchanged UCSC hg38 reference genome, used to extract the long GTEx sequence windows.

The archives are inputs to the benchmark's preparation code. They are not a single training table. Preserve the original filenames when using BEND2_ASSET_DIR.

Atlas

File	Contents
atlas-population.parquet	7,684 rows: 4,223 training, 1,023 validation, 2,411 candidate and 27 excluded sequences. Columns include id, sequence, family_id, population, split, eligible_calibration and program_annotations. All sequences are 145 bases.
atlas-programs.csv	Sixteen activity patterns, with program, on_cells and off_cells. Cell names are separated by |.
atlas-source-activities.parquet	Two measured activities in each of twelve cells for the 5,246 fitting sequences.
atlas-candidate-activities.parquet	The corresponding measurements for the 2,411 candidates, kept separately from fitting data.

The activity columns start with r1_ and r2_; id joins them to the sequence table. Pandas reads id as the index, while Arrow exposes it as a column. Activities are float64 values calculated from UMI-deduplicated RNA and DNA counts, with a pseudocount of 0.5 and normalization to the median of 35 negative controls. No measurement values were changed for this mirror.

Sequences belonging to families that overlap the candidates are excluded from fitting. The 27 excluded rows are retained in the population table to document that separation. BEND uses 7,657 sequences in total.

Sequence-prior panel

prior-panel.parquet contains 1,536 fixed-strand, 1,200-base GRCh38 windows: 1,024 for baseline fitting and 512 for evaluation, separated by genomic blocks. Columns include sequence, genomic coordinates, reference_base, recovery_position and split. The hidden position is index 600 in zero-based coordinates.

prior-query-seeds.json stores the two masking seeds per evaluation window used by the NTv3 paper-prior adapter. It contains no sequences, hidden bases or measurement outcomes. These reference-genome sequences may have appeared in model pretraining.

Additional Siraj measurements

siraj-numeric-reference.npz preserves exact float64 values for all 340,584 mutation/cell records. Its arrays are record_id_bytes, columns and values; the numeric columns are y_ref, y_alt, delta_y and sigma.

siraj_selective.parquet contains 181,458 mutation and no-edit identifiers with columns ID, truth_hepg2 and truth_k562. These effects average the per-repeat log2((RNA + 0.5) / (DNA + 0.5)) difference from the unmutated parent across five repeats. They are separate from the posterior effect estimates in the Siraj archive. The cell-selective task uses 167,958 eligible substitutions across 127 loci; the table also retains the identifiers needed for preparation.

GTEx

gtex-targets.npz contains truth (19,282 genes × 51 tissues), gene_ids, tissues, and training, validation and test index arrays. Targets are log1p(TPM) from the public GTEx v10 median-expression summary.

gtex-genes.tsv records gene identifiers and GENCODE v39 transcription-start coordinates for reference-forward 131,072-base hg38 windows. It preserves the target matrix's row order.

There are 12,146 training and 1,350 validation genes. The stored test set has 5,786 genes; the four exclusions listed in assets.json leave the same 5,782-gene evaluation set used in the paper. They lack complete predictions in the paper's common model/reference comparison.

These are frozen GTEx v10 files and do not necessarily reflect the most current or accurate data from the GTEx Portal. They contain public aggregate expression only. No individual donor records or genotypes are included.

Provenance and checksums

• assets.json is an unchanged copy of the benchmark's source manifest. Its distribution: bundled entries describe how the Python package ships those assets, rather than their location in this Hugging Face repository.
• checksums.json lists the local data/ paths, byte counts and SHA-256 values for this release. The original MD5 checksums remain in assets.json where used by the benchmark.
• LICENSE.md maps files to their source terms and acknowledgements.

The checksum manifest identifies file contents independently of the download host. Rehosting does not change candidate sets, fold assignments, measurements or scoring.