TimeOmni-VL/TSUMM-Suite_Training

🧱 TSUMM-Suite Training Set

This repository provides the training set of TSUMM-Suite, a multimodal dataset for unified time series understanding and generation. It contains forecasting and imputation tasks for TS-image generation. By aligning understanding tasks with generation tasks, TSUMM-Suite enables temporal understanding to improve time series generation.

🎨 Task Illustration

TSUMM-Suite covers two task families:

Generation tasks

• Forecasting. Complete the masked future region of a TS-image and decode the generated image region back into future numerical values.
• Imputation. Complete masked internal regions of a TS-image and decode the generated image region back into missing numerical values.

Understanding tasks

• QA1: Variable Counting. Count the number of variables encoded in a TS-image.
• QA2: Variable Y-Range. Locate the vertical range of a specific variable.
• QA3: Cycle Bounding Box. Locate the spatial region of a target cycle.
• QA4: Mean Comparison. Compare average values across cycles.
• QA5: Anomaly Detection. Identify and locate anomalous regions.
• QA6: Trend Analysis. Analyze temporal patterns within a specified region.

📊 Dataset Information

1. Dataset Components

Component	Task Family	Size	Description
Forecasting	Generation	40,000	TS-image completion samples for future sequence generation
Imputation	Generation	40,000	TS-image completion samples for missing-value reconstruction
TS-image QA	Understanding	9,409	CoT-guided QA pairs for layout-level and signal-level TS-image understanding
TSR-Suite	Reasoning	2,339	CoT-guided temporal reasoning samples used to strengthen temporal reasoning

* Note: This release open-sources the forecasting and imputation generation data only. We do not open-source the TS-image QA (Understanding) and TSR-Suite (Reasoning) training data.

All generation samples are stored in a single train split (80,000 rows total). Each TS-image has resolution 896×896. In the paper, the main TimeOmni-VL experiments use 5k samples per generation task for training.

2. Data Fields

Generation samples contain fields such as:

• sample_id: unique sample identifier
• task: forecasting or imputation
• freq: time series frequency (e.g., H, D, 5T)
• nvars: number of variables
• context_len: context length
• pred_len: prediction or imputation length
• periodicity: periodicity of the series
• input_image: masked TS-image input (896×896)
• target_image: ground-truth completed TS-image (896×896)
• metadata_json: JSON string with normalization, patch layout, mask ratio, and related metadata

🚀 Usage

from datasets import load_dataset

dataset = load_dataset("TimeOmni-VL/TSUMM-Suite_Training")
train = dataset["train"]

forecasting = train.filter(lambda x: x["task"] == "forecasting")
imputation = train.filter(lambda x: x["task"] == "imputation")

print(f"Total: {len(train)}, Forecasting: {len(forecasting)}, Imputation: {len(imputation)}")

📐 Evaluation

For forecasting and imputation, we report normalized Mean Absolute Scaled Error (nMASE). Lower values indicate better numerical generation quality.

For TS-image understanding, we report normalized task-specific scores in the range of [0, 1]. Higher values indicate better understanding performance.

For TSR-Suite reasoning tasks, we report Accuracy (ACC) for text-output tasks and Mean Absolute Error (MAE) for sequence-output tasks.

License

TSUMM-Suite is released under the Apache 2.0 license.

✍️ Citation

@article{guan2026timeomni,
  title={TimeOmni-VL: Unified Models for Time Series Understanding and Generation},
  author={Guan, Tong and Pan, Sheng and Barthelemy, Johan and Li, Zhao and Cai, Yujun and Alippi, Cesare and Jin, Ming and Pan, Shirui},
  journal={arXiv preprint arXiv:2602.17149},
  year={2026}
}