~1M valid SELFIES with seq_len<=25 using FastChemTokenizerSelfies, built and curated from ChemBL34 (Zdrazil et al. 2023), COCONUTDB (Sorokina et al. 2021), and Supernatural3 (Gallo et al. 2023) dataset.
The original dataset was processed by filtering max_seq_len<=25 for training ChemMiniQ3-SAbRLo. The cleaned dataset was then split into 6 approximately equal chunks for training in limited compute.
Curated by: gbyuvd… See the full description on the dataset page:
https://huggingface.co/datasets/gbyuvd/sabrlo-chem-selfies-training.
