---
license: mit
task_categories:
- text-classification
language:
- en
tags:
- membership-inference
- privacy
- machine-learning-security
size_categories:
- n<1K
---

# WIKITEXT Calibration Set for Membership Inference Attacks

This dataset contains a calibration set for training membership inference attack (MIA) classifiers.

## Dataset Information

- **Total Samples**: 500
- **Members**: 250 (sampled from training data)
- **Non-members**: 250 (sampled from unseen test split)
- **Balance**: 50.0% members / 50.0% non-members

## Data Sources

### Members (label=1)
- **Source**: 10K training data (excluding eval set)
- **Dataset**: mia-llm/wikitext-raw-MIA (10K training samples)
- **Filtering**: Excluded samples that appear in evaluation set

### Non-members (label=0)
- **Source**: Unseen test split
- **Original Dataset**: 
  - AG News: `fancyzhx/ag_news` (test split)
  - XSum: `EdinburghNLP/xsum` (test split)
  - WikiText: `Salesforce/wikitext` wikitext-2-raw-v1 (test split)

## Usage

### Load Dataset

```python
from datasets import load_dataset

# Load calibration set
dataset = load_dataset("h0ssn/wikitext-calibration-mia")

# Access data
calibration_data = dataset['train']

print(f"Total samples: {len(calibration_data)}")
print(f"First sample: {calibration_data[0]}")
```

### Train MIA Classifier

```python
from sklearn.linear_model import LogisticRegression
from sklearn.preprocessing import StandardScaler
import numpy as np

# Load calibration set
calibration_data = load_dataset("h0ssn/wikitext-calibration-mia")['train']

# Extract features (example: using loss and gradient)
# You need to compute these features using your target and reference models
features = extract_features(calibration_data['text'])  # Your feature extraction
labels = calibration_data['label']

# Train classifier
scaler = StandardScaler()
features_normalized = scaler.fit_transform(features)

classifier = LogisticRegression(max_iter=1000, random_state=42)
classifier.fit(features_normalized, labels)

# Now use this classifier to predict membership on unknown samples
```

## Data Fields

- `text` (string): The text content
- `label` (int): 1 for members, 0 for non-members
- `source` (string): Origin of the sample ('10k_train' or 'unseen')

## Verification

✅ **No overlap between members and non-members**: All samples have been verified to ensure no duplicates exist between the member and non-member sets.

## Citation

If you use this calibration set in your research, please cite:

```bibtex
@dataset{wikitext_calibration_mia,
  title = {WIKITEXT Calibration Set for Membership Inference Attacks},
  author = {h0ssn},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/h0ssn/wikitext-calibration-mia}
}
```

## Related Datasets

- Evaluation Set: `mia-llm/wikitext_benchmark_roya` (700 samples)
- Training Set: `mia-llm/wikitext-raw-MIA` (10K samples)

## License

MIT License - Free to use for research and commercial purposes.

## Notes

- This calibration set is designed to be used SEPARATELY from the evaluation set
- No samples from this calibration set appear in the evaluation set
- Recommended split: Use this for training MIA classifiers, use evaluation set for testing
- Random seed: 42 (for reproducibility)

## Contact

For questions or issues, please open an issue on the dataset repository.