🔒 safesplit

Leakage-safe train/test splits — group-aware, time-aware, de-duplicated.

sklearn.train_test_split shuffles rows at random. For a lot of real data that quietly leaks: the same patient lands on both sides, a time series trains on its own future, or duplicate rows sit in train and test. Your score looks great and isn't real.

safesplit gives you one honest split:

pip install safesplit

from safesplit import safe_split

# a group is never on both sides
train, test = safe_split(df, group_col="patient_id")

# train on the past, test on the future
train, test = safe_split(df, time_col="date")

# keep near-homologous protein/DNA sequences together
train, test = safe_split(df, seq_col="sequence")

# plain random split, but exact duplicates are dropped first
train, test = safe_split(df, test_size=0.2)

What it does

The companion to LeakHound

safesplit prevents the leaks that LeakHound detects. Split with one, verify with the other:

from safesplit import safe_split
from leakhound import audit           # pip install leakhound-ml

train, test = safe_split(df, group_col="patient_id")
assert audit(train, test, group_col="patient_id").clean

License

GNU AGPL-3.0-or-later © 2026 Joel Gomes. A commercial license is available if the copyleft terms don't fit your use — open an issue.