CoPeP Continual Protein Dataset

This dataset is organized for continual-learning experiments on protein sequences.

Repository

• Dataset repo id: chandar-lab/CoPeP

File layout

• train/: 252 parquet shards (data-00000-of-00252.parquet ... data-00251-of-00252.parquet)
• splits/: 10 task index parquet files (task_0.parquet ... task_9.parquet)
• val/: validation parquet (val.parquet)

Task file mapping

• task_0.parquet -> task_0 (2015)
• task_1.parquet -> task_1 (2016)
• task_2.parquet -> task_2 (2017)
• task_3.parquet -> task_3 (2018)
• task_4.parquet -> task_4 (2019)
• task_5.parquet -> task_5 (2020)
• task_6.parquet -> task_6 (2021)
• task_7.parquet -> task_7 (2022)
• task_8.parquet -> task_8 (2023)
• task_9.parquet -> task_9 (2024)

Important note on splits/task_*.parquet

The splits/task_*.parquet files are index-style split definitions keyed by row_idx. They are intended to be joined with records from train/ (or other source files) using row_idx, rather than treated as standalone full-example datasets.

Migration note

Task index files are now exposed through name="task_splits". The old load_dataset(repo_id, split="task_0") pattern is deprecated.

Basic code: load one task index split

from datasets import load_dataset

repo_id = "__HF_DATASET_REPO__"

# 1) Load train split directly (map-style dataset)
train_ds = load_dataset(repo_id, split="train")

# 2) Load one task index split via the task_splits config.
#    Use streaming=True to avoid Arrow cache materialization for index files.
task0_idx = load_dataset(
    repo_id,
    name="task_splits",
    split="task_0",
    streaming=True,
)

# 3) Materialize examples by selecting train rows using row_idx
task0_rows = [example["row_idx"] for example in task0_idx]
task0_examples = train_ds.select(task0_rows)

print(task0_examples)
print(task0_examples[0])