Cross-Domain ML task type identification

# Dataset Description This repository contains a benchmark dataset collection for machine learning task-type identification. The collection is designed to support the automatic identification of both the data domain and the downstream prediction task of a dataset when the target feature is known. The main benchmark consists of **625 public datasets** from heterogeneous sources, including OpenML, the UCI Machine Learning Repository, Kaggle, and the Time Series Classification Archive. The main dataset covers **tabular** and **time-series** data and is organized according to a two-level taxonomy. In addition, optional NLP datasets are provided separately. These NLP datasets can be downloaded and added to `dataset.zip` as an additional domain. The NLP part contains binary classification, multiclass classification, and text generation tasks. However, this part is currently less curated and less standardized than the tabular and time-series benchmark. ## Dataset Archives The dataset release consists of two main parts: - `dataset.zip`: contains the original dataset files organized in a taxonomy-based folder structure.- `meta_dataset.zip`: contains tabular metadata describing each dataset, including the dataset name, target feature, task label, file location, and additional dataset-level information. Optional additional data: - NLP datasets can be downloaded separately and added to `dataset.zip` as an additional domain.- The NLP data are not yet curated to the same level as the tabular and time-series data and should therefore be considered an additional, preliminary extension. ## Folder Structure The datasets in `dataset.zip` are organized according to their data domain and downstream machine learning task. ```textdataset/├── Time_Series/│ ├── binary_classification/│ ├── multiclass_classification/│ └── regression_forecasting/│├── Tabular/│ ├── binary_classification/│ ├── multiclass_classification/│ └── regression/│└── NLP/ # optional additional domain ├── binary_classification/ ├── multiclass_classification/ └── text_generation/ ## Dataset Distribution The main benchmark contains **625 tabular and time-series datasets**. | Domain | Binary Classification | Multiclass Classification | Regression / Forecasting | Total ||---|---:|---:|---:|---:|| Time Series | 61 | 137 | 101 | 299 || Tabular | 86 | 100 | 140 | 326 || **Total** | **147** | **237** | **241** | **625** | The optional NLP data are provided separately and are **not included** in the main 625-dataset count. ## Metadata Description The `meta_dataset.zip` archive contains a tabular metadata file with one entry per dataset. The metadata table links each dataset to its corresponding file in `dataset.zip` and provides the information required to identify, locate, and load each dataset. The metadata includes the following fields: | Field | Description || `dataset_name` | Name of the dataset. || `data_domain` | Data domain of the dataset, for example `Tabular`, `Time_Series`, or optionally `NLP`. || `downstream_task` | Assigned prediction task, such as binary classification, multiclass classification, regression, forecasting/regression, or text generation. || `target_feature` | Name of the target variable used as the prediction target. || `file_location` | Relative path to the dataset file inside `dataset.zip`. || `dataset_description` | Textual description of the dataset, if available. || `multi_target` | Indicates whether the dataset contains multiple target variables. || `number_of_instances` | Number of samples or records in the dataset. || `number_of_features` | Number of input features, if applicable. || `original_download_link` | Source URL from which the dataset was obtained. | The metadata file can be used to: 1. locate each dataset inside `dataset.zip`,2. identify the corresponding target feature,3. determine the assigned domain and downstream task label,4. access additional information required for loading or processing. ## Dataset Types ### Tabular Classification Tabular classification datasets consist of independent samples represented by feature columns and an explicitly specified target feature. The target feature is categorical. A dataset is assigned to: - **binary classification** if the target contains exactly two classes.- **multiclass classification** if the target contains more than two classes. ### Tabular Regression Tabular regression datasets consist of independent samples represented by feature columns. In contrast to classification, the target feature is continuous and numerical. The objective is to predict a real-valued response from the available input features. ### Time-Series Classification Time-series classification datasets contain observations with an intrinsic temporal or sequential structure. Each sample may be represented by a univariate or multivariate sequence. The target feature is categorical, and the objective is to assign each time-series sample to one of the predefined classes. ### Time-Series Forecasting and Regression Time-series forecasting and regression datasets contain sequential observations and a continuous target feature. In forecasting tasks, the target usually represents a future value estimated from historical observations. In time-series regression tasks, the target may instead describe a continuous property associated with the sequence. Both cases are grouped together because they require the prediction of continuous values from time-dependent data. ### Optional NLP Datasets Optional NLP datasets are provided separately and can be added as an additional domain in `dataset.zip`. They include: - binary text classification,- multiclass text classification,- text generation. The NLP extension is currently less curated than the tabular and time-series parts of the benchmark. In particular, the structure, formatting, and metadata consistency may be less standardized. It should therefore be treated as an additional extension rather than as part of the main cleaned benchmark. ## File Formats The dataset collection preserves the original dataset formats where possible. The included datasets may appear in different file formats, including: - CSV- XLS / XLSX- ARFF- Arrow- JSON This heterogeneity is intentional. The collection is designed to reflect realistic dataset conditions rather than a heavily standardized benchmark format.