READ-Bench

A benchmark for historical-case retrieval in time-series diagnosis: retrieve past cases that share a query's fault / event type, not its signal shape. 12 multivariate datasets across five domains, each with explicit fault-level relevance judgments. See the paper for details (arXiv:2609.32123).

• 🏆 Leaderboard: c3aiia3c/read-bench-leaderboard
• 📥 Submit a method: c3aiia3c/read-bench-results
• 📖 Usage & config: USAGE.md — loading, the tweakable BenchmarkConfig knobs, and reproducing the paper's pollution-level compositions.

Datasets

Dataset	Domain	C	T	Classes	|C| (0/10/20%)	|Q|	rel/q	Imb.
petrobras_3w	Industrial / control	2-7	256	9	1000/1111/1250	100	146.8	23.7
tennessee_eastman	Industrial / control	52	256	20	1000/1111/1250	100	50.0	1.0
damadics	Industrial / control	32	256	4	1000/1111/1250	18	308.8	83.9
voraus	Industrial / control	130	1024	12	528/587/660	100	68.9	15.6
exathlon	Cyber-physical & IT	51	256	6	1000/1111/1250	100	215.9	13.2
telecom_ts	Cyber-physical & IT	18	128	11	977/1086/1221	100	117.8	5.3
hai	Cyber-physical & IT	79	256	6	305/339/381	28	60.8	10.0
rcaeval	Cyber-physical & IT	429	256	5	700/778/875	50	140.0	1.0
road	Automotive	664	256	6	71/79/89	25	22.4	37.0
mit_bih	Healthcare	2	256	4	1000/1111/1250	100	336.9	10.7
ctsr	General / cross-domain	1	512	94	1000/1111/1250	100	10.6	1.2
rats40k	General / cross-domain	1-9	16-128	20	1000/1111/1250	100	151.2	342.0

_{C channels, T window length, |C| corpus size at pollution levels 0/10/20%, |Q| queries, rel/q mean same-type items per query, Imb. class-imbalance ratio (paper Table 3).}

Usage

from datasets import load_dataset
ds = load_dataset("c3aiia3c/read-bench", "ctsr", split="train")
import torch; ts = torch.tensor(ds[0]["ts_values"])   # (T, C) window

Each row is one window: ts_values ([T][C] floats), channels, label (Normal/Anomaly), class_label (the retrieval target), rca_target, rca_affected, dataset, entry_id, native_split, extras (JSON).

Citation

@article{readbench2026,
  title  = {READ-Bench: Benchmarking Historical Instance Retrieval for Time-Series Diagnosis},
  author = {Pastrana, Gerardo and Li, Haojun and Mehta, Dhruv and Vyas, Anoushka and
            Khoshfetrat Pakazad, Sina and Ohlsson, Henrik and Paparrizos, John},
  journal = {arXiv preprint arXiv:2609.32123},
  year    = {2026}
}

Each config derives from its original source dataset; please also cite the underlying datasets you use.