ReTabAD: Benchmark for LLM-based Tabular Anomaly Detection [ICLR 2026]
π Paper | π Project Page | π» GitHub
Official data for ReTabAD: A Benchmark for Restoring Semantic Context in Tabular Anomaly Detection (ICLR 2026).
Existing tabular anomaly detection benchmarks strip away semantic context by converting textual features into opaque encodings and discarding descriptive metadata. ReTabAD restores this context: each of the 20 datasets comes with structured textual metadata, so that methods, including large language models (LLMs), can use the domain knowledge experts rely on. We hope ReTabAD serves as a common testbed for researchers exploring LLMs for tabular anomaly detection.
Structure
Each dataset directory contains:
{dataset}.csv: cleaned datapoints ready for analysis{dataset}.json: semantic metadata (column descriptions, logical types, and characterizations of normal data)preprocess.py: data preparation script (filtering NA, normalization, etc.)
Datasets
20 datasets with varying characteristics:
- Size: 159 to 50,000 datapoints
- Features: 6 to 42 columns
- Anomaly ratio: 0.4% to 33.3%
See dataset_statistics.csv for the complete overview.
Usage
huggingface-cli login
huggingface-cli download LGAI-DILab/ReTabAD --repo-type dataset --local-dir ./data
import json
import pandas as pd
name = "wine"
data = pd.read_csv(f"data/final_data/{name}/{name}.csv")
with open(f"data/final_data/{name}/{name}.json") as f:
metadata = json.load(f)
To run classical, deep learning, and LLM-based detectors on ReTabAD, see the GitHub repository.
Citation
@inproceedings{yoon2026retabad,
title = {{ReTabAD}: A Benchmark for Restoring Semantic Context in Tabular Anomaly Detection},
author = {Sanghyu Yoon and Dongmin Kim and Suhee Yoon and Ye Seul Sim and Seungdong Yoa and Hye-Seung Cho and Soonyoung Lee and Hankook Lee and Woohyung Lim},
booktitle = {The Fourteenth International Conference on Learning Representations},
year = {2026},
url = {https://openreview.net/forum?id=UFwgg44VZq}
}
- Downloads last month
- 228