---
license: mit
task_categories:
  - tabular-classification
  - tabular-regression
language:
  - en
size_categories:
  - n<1K
  - 1K<n<10K
tags:
  - imputation
  - missing-data
  - tabular
pretty_name: imputify datasets
configs:
  - config_name: iris
    data_files: iris.parquet
  - config_name: wine
    data_files: wine.parquet
  - config_name: diabetes
    data_files: diabetes.parquet
  - config_name: breast_cancer
    data_files: breast_cancer.parquet
  - config_name: titanic
    data_files: titanic.parquet
  - config_name: heart_disease
    data_files: heart_disease.parquet
  - config_name: blood_transfusion
    data_files: blood_transfusion.parquet
  - config_name: thalassemia
    data_files: thalassemia.parquet
  - config_name: ilpd
    data_files: ilpd.parquet
  - config_name: spas_agri
    data_files: spas_agri.parquet
---

# imputify datasets

Curated tabular datasets bundled with the [`imputify`](https://github.com/gabfssilva/imputify)
library for examples, tests, and missing-data benchmarks. Every file is a single parquet:
feature columns first, target column last, no missing values. Column names are
`snake_case`; columns that the upstream sources include for identification but not
modeling (IDs, free-text names) have been dropped.

## Usage

```python
from imputify import load, introduce_missing

X, y = load("iris")                              # clean
X_missing, mask = introduce_missing(X, 0.3)      # ampute for experiments
```

`load` caches downloads through `huggingface_hub` (default `~/.cache/huggingface/hub`).
The complete catalogue is exposed as `imputify.DATASETS` and the literal type as
`imputify.Dataset`.

## Catalogue

| Name | Rows | Features | Numeric | Categorical | Target dtype | Original source |
|------|-----:|---------:|--------:|------------:|--------------|-----------------|
| `iris`              |   150 |  4 |  4 | 0 | `int64`    | `sklearn.datasets.load_iris` (Fisher, 1936) |
| `wine`              |   178 | 13 | 13 | 0 | `int64`    | `sklearn.datasets.load_wine` (UCI Wine Recognition) |
| `diabetes`          |   442 | 10 | 10 | 0 | `float64`  | `sklearn.datasets.load_diabetes` (Efron et al., 2004) |
| `breast_cancer`     |   569 | 30 | 30 | 0 | `int64`    | `sklearn.datasets.load_breast_cancer` (UCI WDBC) |
| `titanic`           | 1 043 |  7 |  5 | 2 | `category` | OpenML `titanic` |
| `heart_disease`     |   270 | 13 | 13 | 0 | `category` | UCI Heart Disease / OpenML `heart-statlog` |
| `blood_transfusion` |   748 |  4 |  4 | 0 | `category` | UCI Blood Transfusion Service Center |
| `thalassemia`       |   606 | 18 | 12 | 6 | `object`   | [Mendeley Data `8kcdkxmcjw`](https://data.mendeley.com/datasets/8kcdkxmcjw/1) — Pabna, Bangladesh thalassemia cohort, Data in Brief 2025 |
| `ilpd`              |   583 | 10 |  9 | 1 | `category` | UCI Indian Liver Patient Dataset (ILPD) |
| `spas_agri`         | 4 191 | 14 |  9 | 5 | `object`   | [Mendeley Data `cphdw4z5kw`](https://data.mendeley.com/datasets/cphdw4z5kw/2) — SPAS-Dataset-BD, Bangladesh precision agriculture, Data in Brief 2025 |

Counts come from running the loader once: rows = `len(X)`, features = `X.shape[1]`,
numeric/categorical split = `select_dtypes(...)`.

## Format

Each parquet has the **target** as the last column. `imputify.load(name)` splits it back
into `(X, y)`:

```python
df = pd.read_parquet("iris.parquet")
X, y = df.drop(columns=["target"]), df["target"]
```

Categorical features are stored as pandas `category` (titanic) or `object`
(thalassemia, spas_agri) and round-trip through parquet without manual casting.

## Licensing & attribution

Datasets carry their original licences. `iris`, `wine`, `diabetes`, and `breast_cancer`
ship with scikit-learn (BSD-compatible). The UCI / OpenML datasets are redistributed
under their respective terms. The `thalassemia` and `spas_agri` derivatives are from
Mendeley Data (CC BY 4.0); cite the Data in Brief 2025 papers in publications.