README.md

# HPO-B
HPO-B is a benchmark for assessing the performance of black-box HPO algorithms. This repo contains the code for easing the consumption of the meta-dataset and speeding up the testing. 

## Meta-Dataset 

The meta-dataset contains evaluations of the accuracy for different search-spaces on different datasets. For more details on the meta-dataset, refer to our [paper](https://arxiv.org/pdf/2106.06257.pdf). It is presented in three versions:

- **HPO-B-v1**: The raw benchmark of all 176 meta-datasets.
- **HPO-B-v2**: Subset of 16 meta-datasets with the most frequent search spaces.
- **HPO-B-v3**: Split of HPO-B-v2 into training, validation and testing. 

## Meta-dataset format

As described in the paper, the meta-dataset follows a JavaScript Object Notation (JSON) to encapsulate the evaluations. In Python, this corresponds to nested dictionaries, where the first level key corresponds to the **search space ID**, the second level key contains the **dataset ID**, and finally the last level contains the list of hyperparameter configurations (**X**) and its response (**y**).


```python
meta_dataset = { "search_space_id_1" : { "dataset_id_1": {"X": [[1,1], [0,2]],
                                                  "y": [[0.9], [0.1]]},
                               { "dataset_id_2": ... },
                 "search_space_id_2" : ...
                                
                }