swhitfield/biomap-research-metal_ion_binding

metal_ion_binding

Sourced from biomap-research/metal_ion_binding and prepared for Hugging Face datasets usage.

Data files

Parquet files are stored under data/ using Hugging Face split naming conventions (train-*, validation-*, test-*).

Preparation

• Preprocess mode: minimal.
• Seed: 1957723.
• No max sequence length filter was applied.
• Renamed source columns: label -> targets, seq -> sequence.
• Columns: id, sequence, targets, split.
• Validation handling: upstream did not provide a validation split, so validation was sampled from train using seed 1957723 and matched to the test split size. MMseqs2 cluster boundaries were used when the mmseqs binary was available, with seeded random row sampling as the fallback.
• Split subsets: canonical train/validation/test preserve the downloaded source splits; _random, _stratified, and _cluster suffixes select pooled alternative partitions generated with seed 1957723. Each pooled partition reuses the canonical split sizes.

Sequence clustering

MMseqs2 (mmseqs easy-linclust) settings used for cluster-based split assignment:

• Minimum sequence identity (--min-seq-id): 0.3
• Minimum coverage (-c): 0.8
• Coverage mode (--cov-mode): 1
• Threads: 1
• Seed (cluster ordering / tie-breaking): 1957723

Split sizes

• test: 1332 rows
• test_cluster: 1332 rows
• test_random: 1332 rows
• test_stratified: 1332 rows
• train: 4668 rows
• train_cluster: 4668 rows
• train_random: 4668 rows
• train_stratified: 4668 rows
• validation: 1332 rows
• validation_cluster: 1332 rows
• validation_random: 1332 rows
• validation_stratified: 1332 rows

Loading

from datasets import load_dataset

ds = load_dataset("<owner>/<repo>")
print(ds)