Datasets:

day3
/

RATs40K-3view

Everything needed to run the three-view diagnosis + residual rescue routing experiments without generating anything: the original RATs40K, a corrected copy of it, the repeat-safe three-fold split, and one rendered image per diagnosis view.

Put it at data/ in the code repository that reads it, then scripts/run/prepare_data.sh both expands the archives and you can run.

File	Size	What
`RATs-Uni-TSImage_Reason.json`	93M	original univariate split, forwarded unchanged
`RATs-Multi-TSImage_Reason.json`	8M	original multivariate split, forwarded unchanged
`figures.tar`	700M	the univariate figures RATs40K ships (36,331)
`figures_multi.tar`	84M	the multivariate figures RATs40K ships (3,274)
`fixed/`	61M	the corrected dataset, plus a manifest of every change
`folds_uni.json`, `folds_multi.json`	0.5M	the repeat-safe three-fold split
`views_uni.tar`	~2.2G	one image per view, univariate (108,900)
`views_multi.tar`	292M	one image per view, multivariate (9,822)
`route_labels_*.json`	0.8M	route-token training labels

What is ours and what is forwarded

The two RATs-*-TSImage_Reason.json files and the two figures*.tar are RATs40K as released, repackaged from zip to tar so one unpack command covers the whole bundle. Everything else was produced by us:

fixed/ — a corrected copy. Two defects were found by checking the release against the Time-RA paper:

Univariate: the test split is contaminated. 3,790 of its samples carry multivariate class names, and their ActionIDs follow the multivariate convention exactly (verified on 4,847/4,847 including Normal). They are moved into a separate split rather than deleted, so nothing is lost and the original is recoverable. The train split is correct and is passed through byte for byte.
Multivariate: the data is correct — ids 0-6 match Table 7 of the paper. The prompt in the original code is not: it invents an eighth class, "Local Structural Jump", and shifts Joint Space, Principal Component Space and Collinearity Shift from 4/5/6 to 5/6/7. So the multivariate file is copied through unchanged and the corrected class list ships beside it as anomaly_dict_multi_fixed.py.

fixed/manifest.json records every change with hashes.

folds_*.json — a three-fold split of the training set that keeps near-duplicate series together, so an out-of-fold prediction is never a lookup of its own near-twin. Seeded; regenerating gives the same partition.

views_*.tar — three images per sample:

View	Image
`base`	the dataset's own figure, copied unchanged
`msd`	the series under two mean-pooled resolutions; multivariate also reorders channels by correlation, bands synchronous jumps, and adds two correlation heatmaps
`key`	the candidate anomalous regions, numbered so the marks and the text digest name the same indices

base is deliberately not redrawn: it is the control, and redrawing it would add rendering differences (axis margins, line width, legend size) to the one view that carries no new evidence.

For msd and key, whether each channel is z-scored is decided per sample: it happens only when the median channel occupies less than 5% of the shared axis, which is where the smallest channels would otherwise be zero pixels tall. That quantity is bimodal on this data — squashed samples sit near 0.01, readable ones near 0.70, with nothing in between — so the threshold falls in the gap and any value from 0.02 to 0.20 changes three samples.

Sizes and layout after unpacking

data/
├── RATs-Uni-TSImage_Reason.json
├── RATs-Multi-TSImage_Reason.json
├── fixed/{RATs-Uni,RATs-Multi}-TSImage_Reason.json, manifest.json, anomaly_dict_multi_fixed.py
├── folds_uni.json, folds_multi.json
├── route_labels_qwen2.5-7b-route3.json, route_labels_empty.json
├── figures/{train,test}/            36,331 images
├── figures_multi/{train,test}/       3,274 images
└── views/
    ├── uni/{TSAD_train,TSAD_test}/  108,900 images
    └── multi/{TSAD_train,TSAD_test}/  9,822 images

About 3.4 GB compressed, 4.1 GB unpacked.

Archives rather than loose files because the view sets alone are 118,722 small JPEGs; the original RATs40K ships figures.zip for the same reason.

Which split each table uses

Not a free choice — it is what produced the published numbers:

Univariate reads the original split. Its test set holds 6,034 scorable samples; the corrected one is a 2,244-sample subset of those, so one prediction file is scored against both.
Multivariate reads the corrected split, which for multivariate is byte-identical to the original — the correction there is to the class list, not the data.

TSAD_train is byte-identical between the two in both modalities, so training and the fold manifest are unaffected by the choice.

Citation

Cite the original dataset:

@article{yang2026timera,
  title={TIME-RA: Towards Time Series Reasoning for Anomaly Diagnosis with LLM Feedback},
  author={Yang, Yiyuan and Liu, Zichuan and Song, Lei and Ying, Kai and Wang, Zhiguang and Bamford, Tom and Vyetrenko, Svitlana and Bian, Jiang and Wen, Qingsong},
  journal={arXiv preprint arXiv:2507.15066},
  year={2025}
}