ChronoSteer-100K

This dataset provides the pre-training and fine-tuning resources used for ChronoSteer, a multimodal time series forecasting framework that aligns textual revision instructions with time series foundation models.

ChronoSteer-100K is designed to learn controllable forecasting behaviors from synthetic cross-modal supervision. It associates historical time series with a compact codebook of revision instructions and their corresponding forecast targets.

Dataset Construction

The time series are collected from the Monash Time Series Forecasting Repository. A sliding window of length 160 and stride 32 is applied to produce approximately 15 million normalized slices. Each slice is divided into:

• a historical series of length 128;
• a future series of length 32.

K-means clustering is used to select 100,000 representative slices. For each slice, Chronos-Bolt-base first produces an initial forecast. Nine transformation functions are then applied to construct synthetic targets corresponding to the following revision instructions:

1. Keep Unchanged
2. Increase Trend
3. Reduce Trend
4. Expand Amplitude
5. Compress Amplitude
6. Elevate Peaks
7. Lower Peaks
8. Raise Troughs
9. Deepen Troughs

This process produces 900,000 instruction-series pairs for synthetic cross-modal pre-training. The observed future series is also included for pseudo-label-guided fine-tuning.

Contents

The repository contains:

ChronoSteer-100K/
├── data/
│   └── ChronoSteer-100K.json
└── synth.ipynb

Each sample in ChronoSteer-100K.json contains:

Field	Description
hist_series	Normalized historical time series with length 128.
true_series	Normalized observed future series with length 32.
pred_series	A dictionary containing the synthetic forecast target for each of the nine revision instructions.

The synth.ipynb notebook provides the synthetic transformation procedure used to construct the instruction-specific forecast targets.

For model configuration, training commands, and evaluation instructions, please refer to the ChronoSteer GitHub repository.

Citation

If you use this resource, please cite ChronoSteer:

@article{chronosteer,
    author  = {Chengsen Wang and Qi Qi and Zhongwen Rao and Lujia Pan and Jingyu Wang},
    title   = {ChronoSteer: Bridging Large Language Model and Time Series Foundation Model via Synthetic Cross-Modal Alignment Dataset},
    journal = {ACM Transactions on Knowledge Discovery from Data},
    year    = {2026},
}