Top-level folder: raw_ghz_a.zip/ raw_test_ghz_a.zip/ – contains the test subset for the ghz_a search space and includes raw_test_ghz_a_data.zip with .pckl files storing each PQC along with its metadata and test_ghz_a.db file - circuits metadata as DB raw_train_ghz_a.zip/ – contains the train subset without augmentation for the ghz_a search space and includes raw_train_ghz_a_data.zip with .pckl files storing each PQC along with its metadata and train_ghz_a.db file - circuits metadata as DB raw_train_ghz_a_augmented.zip/ – contains the train subset incl. augmented PQCs for the ghz_a search space and includes raw_train_ghz_a_augmented_data.zip with .pckl files storing each PQC along with its metadata and train_ghz_a_augmented.db file -circuits metadata as DV Top-level folders: raw_ghz_b.zip/ has the same structure but data for the search space ghz_b Top-level folders: ls_a.zip/ raw_ls_a_data.zip/ – contains .pckl files storing each PQC along with its metadata for the search space l ls_a.db file - circuits metadata as DB Note that for train/test splitting, ls_a space uses the automatical split specified in gen_dataset.py with random seed `42`. Top-level folders: graph_data_ghz_a.zip/, graph_data_ghz_b.zip/,graph_data_ls_a.zip/ – contain .pt files, i.e., all data subsets for particular search spaces, with circuits converted into directed acyclic graph respresentation (DAG). This representation can be directly used to tran a GCN. 

This dataset supports the SQuASH benchmark for Quantum Architecture Search (QAS), as presented in our paper. It includes training and evaluation data used for surrogate model learning, structured into multiple problem instances. Each subdirectory contains a database file with information extracted from .pkl files, such as initial PQC, optimal PQC and target evaluation metric, e.g., fidelity or train/test accuracy. The dataset is organized for direct integration with the SQuASH GitHub repository and is designed to accelerate QAS research and support reproducible benchmarking.