VeriTime: Time Series Reasoning via Process-Verifiable Thinking Data Synthesis and Scheduling for Tailored LLM Reasoning

This is the dataset associated with our paper:

Time Series Reasoning via Process-Verifiable Thinking Data Synthesis and Scheduling for Tailored LLM Reasoning

Jiahui Zhou, Dan Li, Boxin Li, Xiao Zhang, Erli Meng, Lin Li, Zhuomin Chen, Jian Lou, See-Kiong Ng

ICML 2026  |  Paper  

Dataset Construction Pipeline: TSRgen

TSRgen is an automated pipeline that constructs a time series–text multimodal reasoning dataset with process-verifiable annotations. Our dataset provides step-level verifiable labels at intermediate reasoning steps, enabling fine-grained RL supervision over the full CoT trajectory.

Pipeline overview:

1. Data Source Selection — Synthetic and real-world time series, 7 tasks across 4 domains.
2. Task Filtering & Categorization — Rule-based extraction into scenario-based and knowledge-based tasks; true/false, multiple-choice, and open-ended Q&A formats.
3. TS-Tailored 6-Step CoT — Each sample includes a structured thinking process: ① task intent analysis → ② key pattern selection → ③ segment-level TS analysis → ④ preliminary answer → ⑤ self-reflection → ⑥ final summarization.
4. Quality Control — DeepSeek-R1 generates reasoning trajectories; answer-correctness filtering + cross-LLM validation.
5. Verifiable Annotation Extraction — Process-level labels derived at Steps 1, 2, 4, 6 for independent verification.

Why Process-Verifiable?

Our dataset provides step-level verifiable labels, multi-objective reward signals, and a structured TS-tailored CoT that cuts token usage while improving accuracy. This enables compact 3B–4B models to match or outperform frontier LLMs.

Dataset Statistics

Category	Task	#Samples	Avg. Time Points	Avg. Token Count
Scenario-based	Anomaly Detection	1,180	300	2,759
	Scenario Attribution	930	300	2,932
	Inferential Calculation	410	324	2,783
Knowledge-based	CTU	270	720	5,942
	ECG	780	500	4,239
	EMG	450	600	4,990
	RCW	320	500	4,201
Total		4,340

• Scenario-based tasks use synthetic time series generated under predefined conditions, with a training-to-test ratio of approximately 5:1.
• Knowledge-based tasks are derived from real-world professional domain datasets with original train/test splits preserved.

Data Format

Each sample is stored as a JSON object in JSONL format with the following fields:

Field	Type	Description
task	string	Task type (e.g., "Anomaly detection", "Scenario attribution", "RCW")
question	string	Full question with context, metric description, and options
output	string	Model output with answer and explanation
label	string	Ground-truth answer label
cot	string	Complete Chain-of-Thought reasoning (with <THINK> and <ANSWER> tags)
step1_label	string	Verifiable label for Step 1 (task intent)
step2_label	string	Verifiable label for Step 2 (key patterns)
step4_label	string	Verifiable label for Step 4 (preliminary answer)
step6_label	string	Verifiable label for Step 6 (final answer)
timeseries2	list	Time series data (list of float arrays)

File Structure

Dataset/
├── Scenario/
│   ├── scenario_shuffled_train.jsonl   # Scenario-based training set
│   └── scenario_shuffled_test.jsonl    # Scenario-based test set
└── Knowledge/
    ├── knowldege_shuffled_train.jsonl  # Knowledge-based training set
    └── knowledge_test.jsonl            # Knowledge-based test set

Citation

@inproceedings{zhou2026veritime,
  title={Time Series Reasoning via Process-Verifiable Thinking Data Synthesis and Scheduling for Tailored {LLM} Reasoning},
  author={Jiahui Zhou and Dan Li and Boxin Li and Xiao Zhang and Erli Meng and Lin Li and Zhuomin Chen and Jian Lou and See-Kiong Ng},
  booktitle={Forty-third International Conference on Machine Learning},
  year={2026}
}

Acknowledgements

The scenario-based portion of this dataset builds upon the ChatTS-Training-Dataset. We thank the ChatTS team for making their data publicly available.

License

This dataset is released under the Apache 2.0 License.