softdata 0.1.1

Soft Tech Talks: data loading, cleaning, and splitting in 1-3 calls.

softdata

softdata is the "data in, ready out" layer of Soft Tech Talks. One-liners to load, clean, and split your dataset with safe defaults and friendly errors.

from softdata import load, clean, split

df = load("iris")                           # or load("csv", path="students.csv")
df = clean(df, impute="median", encode="auto")
Xtr, Xval, Xte, y = split(df, target="species", strategy="stratified")

Install (local)

pip install -e .
# or build: python -m build  (needs `pip install build`)

API

load(source, **kwargs)

• Built-ins: "iris", "wine", "breast_cancer" (scikit-learn toys)
• Files: "csv" (needs path=), "parquet" (needs path=)

clean(df, impute="median", encode="auto", drop_leaky=None, datetime_auto=True)

• Detects numeric/categorical/date columns
• Imputes numeric (median/mean) and categorical (most frequent)
• Encodes categorical columns with one-hot (drop-first) when encode="auto"
• Preserves the original target column (do encoding only on features)

split(df, target, strategy="auto", test_size=0.2, val_size=0.1, random_state=42)

• If strategy="auto", uses stratified split for discrete targets (<= 20 unique values), else random
• Returns X_train, X_val, X_test, y_dict where y_dict has "train"|"val"|"test"

Example

from softdata import load, clean, split
df = load("iris")
df = clean(df)
Xtr, Xval, Xte, y = split(df, target="target")
print(Xtr.shape, Xval.shape, Xte.shape)

Tests

pip install -r requirements-dev.txt
pytest -q