---
license: mit
task_categories:
  - tabular-classification
language:
  - en
tags:
  - label-noise
  - instance-dependent-noise
  - benchmark
  - corruption
  - tabular
  - adult
  - reproducibility
size_categories:
  - 100K<n<1M
pretty_name: "CILN-Bench: Adult"
---

# CILN-Bench: Adult

Instance-dependent label noise benchmarks built from controlled tabular corruptions.

This dataset is the **Adult** component of CILN-Bench. We corrupt the UCI Adult Income *noisy-label-train* (NLT) split with 5 standard tabular corruptions (Jenga family — 3 missingness mechanisms + 2 value perturbations, 3 severities each → 15 settings), then let a 5-voter pool classify the corrupted rows. The resulting soft labels are released as a noisy-label benchmark.

* **Code, examples, and reproducibility tests:** <https://github.com/sh-islam/ciln-bench>
* **Companion datasets:** [ciln-bench-cifar10](https://huggingface.co/datasets/sh-islam/ciln-bench-cifar10), [ciln-bench-mnist](https://huggingface.co/datasets/sh-islam/ciln-bench-mnist)

## Settings released

5 corruption types × 3 severities = **15 settings**.

| Family | Corruptions |
|---|---|
| Missingness | missing_mcar, missing_mar, missing_mnar |
| Value perturbation | gaussian_noise, scaling |

Severity controls the corrupted-row fraction (sev 1 → 5%, sev 3 → 25%, sev 5 → 50%). For `missing_mar`, corruption probability depends on the observed `sex` attribute; for `missing_mnar`, it depends on the values being corrupted.

Noise rate ranges from 14.7% to 26.3% across the 15 settings.

## Voter pool

5 voters: **XGBoost**, **CatBoost**, **RTDL-MLP**, **FT-Transformer**, **TabPFN**.

## Repository layout

```
settings/
├── gaussian_noise_sev1/
│   ├── noisy_label_train/
│   │   ├── adult_corrupted.parquet   # corrupted feature rows
│   │   ├── labels.npy                # (N,) int — ground-truth income label
│   │   ├── softmax_xgboost_dummyna.npy
│   │   ├── softmax_catboost.npy
│   │   ├── softmax_mlp.npy
│   │   ├── softmax_ft_transformer.npy
│   │   ├── softmax_tabpfn.npy
│   │   ├── avg_softmax.npy
│   │   ├── manifest.json
│   │   └── params.jsonl
│   └── noisy_label_valid/
│       └── ... (same structure)
└── ... (15 settings total)
```

## How to load

```python
import numpy as np
import pandas as pd
from huggingface_hub import snapshot_download

local = snapshot_download(
    repo_id="sh-islam/ciln-bench-adult",
    repo_type="dataset",
    allow_patterns=["settings/missing_mar_sev3/noisy_label_train/*"],
)

features = pd.read_parquet(f"{local}/settings/missing_mar_sev3/noisy_label_train/adult_corrupted.parquet")
labels   = np.load(f"{local}/settings/missing_mar_sev3/noisy_label_train/labels.npy")
print(features.shape, labels.shape)
```

`allow_patterns` is a filter that limits which files get downloaded. Pass a glob (or a list of globs) and only matching files come down. Omit it to download the full dataset.

## Citation

```bibtex
@inproceedings{cilnbench2027,
  title  = {CILN-Bench: A Benchmark for Corruption-Induced Label Noise},
  author = {Islam, Shadman and Kristiadi, Agustinus and Milani, Mostafa},
  booktitle = {ICDE},
  year   = {2027}
}
```

## License

MIT.