| --- |
| license: other |
| task_categories: |
| - time-series-forecasting |
| task_ids: |
| - univariate-time-series-forecasting |
| - multivariate-time-series-forecasting |
| annotations_creators: |
| - no-annotation |
| source_datasets: |
| - original |
| tags: |
| - forecasting |
| - benchmark |
| - fev |
| - arxiv:2509.26468 |
| - tsfile |
| - modality:timeseries |
| - timeseries |
| - format:tsfile |
| size_categories: |
| - 10K<n<100K |
| pretty_name: hermes (TsFile format) |
| configs: |
| - config_name: default |
| data_files: |
| - split: train |
| path: "**/*.tsfile" |
| modality: |
| - timeseries |
| --- |
| |
| # hermes (TsFile format) |
|
|
| This repository contains time-series forecasting data stored in [Apache TsFile](https://tsfile.apache.org/) format. |
|
|
| ## Summary |
|
|
| - FEV subset: `hermes` |
| - Unified source collection: [`autogluon/fev_datasets`](https://huggingface.co/datasets/autogluon/fev_datasets) |
| - Original source: https://github.com/etidav/HERMES |
| - Paper / citation: [[12]](https://arxiv.org/abs/2202.03224) |
| - Series: 10,000 |
| - Modalities: Time-series |
| - TsFile rows (flattened observations): 5,220,000 |
| - Frequencies: source-defined |
| - TsFile files: 3 |
| - Time precision: milliseconds (`INT64`). |
|
|
| Licensing and citation requirements follow the original source. This repository does not claim ownership of the original data. |
|
|
| ## Dataset Statistics |
|
|
| | Series | Median series length | TsFile rows (observations) | Dynamic columns | Static columns | Data files | |
| |---:|---:|---:|---:|---:|---| |
| | 10,000 | 261 | 5,220,000 | 2 | 2 | `hermes_1..hermes_3.tsfile` (3 shards) | |
|
|
| ## Files |
|
|
| The Hugging Face dataset card YAML points `configs.data_files` to all `*.tsfile` files in this repository. |
|
|
| - `hermes_1.tsfile` |
| - `hermes_2.tsfile` |
| - `hermes_3.tsfile` |
|
|
| ## TsFile Storage Model |
|
|
| - Each original series (`id`) is stored as one TsFile device. |
| - Static covariate columns are stored as TAG columns: `country, category`. |
| - Time-varying targets and dynamic covariates are stored as FIELD measurements. |
| - Source `timestamp` values are mapped to the TsFile `Time` column as millisecond timestamps. |
| - Table name(s): hermes. |
|
|
| ### Column Schema |
|
|
| | Column | Role | TsFile type | |
| |---|---|---| |
| | `Time` | Time column | INT64 | |
| | `id` | TAG (device dimension) | STRING | |
| | `country` | TAG (device dimension) | STRING | |
| | `category` | TAG (device dimension) | STRING | |
| | `target` | FIELD (measurement) | FLOAT | |
| | `external` | FIELD (measurement) | FLOAT | |
|
|
| ## Conversion Notes |
|
|
| - The source FEV format stores each time series as one nested row containing `id`, `timestamp[]`, and target or covariate arrays. |
| - The TsFile conversion flattens those nested arrays into long rows. Therefore, the `TsFile rows` values above correspond to the number of timestamped observations after flattening. |
| - TAG columns identify the device and static metadata. FIELD columns contain values that change over time. |
| - Large logical tables may be split into multiple `.tsfile` shards such as `<name>_1.tsfile`, `<name>_2.tsfile`, and so on. Shards listed for the same frequency belong to the same logical table. |
|
|
| ## Reading Example |
|
|
| ```python |
| from tsfile import TsFileReader |
| |
| reader = TsFileReader("hermes_1.tsfile") |
| schemas = reader.get_all_table_schemas() |
| # Table name(s): hermes |
| ``` |
|
|
