Evaluation Dataset A and B for ChatTS

• We provide the two evaluation datasets we collected, as mentioned in the paper. Each dataset sample contains the following fields: `timeseries`, `question`, `answer` (standard answers in text format, for reference only), `attributes` (structured labels used for result evaluation), and `ability_types` (the tasks included in the question). Please note that, to reduce the evaluation cost, we have merged different questions for the same time series into a single `question`, using numbering to distinguish between the different questions. Therefore, the actual number of questions in the evaluation dataset may be greater than the number of `timeseries` entries. Additionally, please note that some tasks in inductive reasoning and alignment are grouped into the same question, as the inductive reasoning tasks involve explaining the physical meanings of time series attributes.

• The `MCQ2` dataset is a third-party open-source dataset, we do not provide it in this repository. Please download it directly via https://github.com/behavioral-data/TSandLanguage.

Data Source

Dataset A includes real-world time series data collected from the following sources:

• NAB (https://github.com/numenta/NAB)
• Weather (https://www.bgc-jena.mpg.de/wetter/)
• Oracle (https://zenodo.org/records/6955909)
• AIOps (https://github.com/netmanaiops/kpi-anomaly-detection)