remove_empty/remove_empty.py

from pathlib import Path

import mteb

log_file_path = Path("remove_empty.log")

# remove log file if exists
if log_file_path.exists():
    log_file_path.unlink()

tasks = mteb.get_tasks(tasks=["STS22"])
task = tasks[0]

task.load_data()


def filter_sample(x):
    if len(x["sentence1"]) > 0 and len(x["sentence2"]) > 0:
        return True
    log = f"Filtered: {x['sentence1']} -- {x['sentence2']}"
    with open(log_file_path, "a") as f:
        f.write(log + "\n")
    print(log)
    return False


for hf_subset in task.dataset:
    _ds = task.dataset[hf_subset]
    for split in _ds:
        ds = _ds[split]
        # filter empty sentences
        n_samples = len(ds)
        ds = ds.filter(lambda x: filter_sample(x))
        n_left = len(ds)
        log = f"Filtered {n_samples - n_left} samples from {n_samples} in {hf_subset} - {split}"
        with open(log_file_path, "a") as f:
            f.write(log + "\n")
        print(log)
        _ds[split] = ds

    task.dataset[hf_subset] = _ds

save_path = Path(__file__).parent.parent

for hf_subset in task.dataset:
    _ds = task.dataset[hf_subset]
    for split in _ds:
        ds = _ds[split]
        ds.to_json(save_path / split / (hf_subset + ".jsonl.gz"), compression="gzip")
        log = f"Saved {hf_subset} - {split} to {save_path / split / (hf_subset + '.jsonl.gz')}"

        with open(log_file_path, "a") as f:
            f.write(log + "\n")
        print(log)