TACIT: model checkpoints and USPTO-MIT blockspace data

Model checkpoints and data for TACIT (TAckling Chemistry through ImpliciT negative learning), a reaction outcome prediction model that ranks pairs of chemical bonds against a stop action and is trained on positive reactions only. Code: github.com/GloriusGroup/tacit. TACIT models (trained on the blockspace-filtered USPTO-MIT train split, final checkpoints after 50 epochs, weights only): tacit.pt, tacit.yaml: TACIT, trained with the auxiliary stop loss. tacit_no-stop.pt, tacit_no-stop.yaml: TACITno-stop, trained without the auxiliary stop loss. Data: uspto_mit_blockspace.zip contains the USPTO-MIT train, validation and test splits (Jin et al., 2017) after blockspace processing, with the elementary-step decompositions used for training. A README describes the fields and the filter used for training (334,510 train, 24,522 validation and 32,860 test reactions). Retrained baselines: baseline_megan.zip, baseline_molecular_transformer.zip, baseline_localtransform.zip and baseline_reactiont5.zip contain MEGAN, the Molecular Transformer, LocalTransform and ReactionT5, each trained with the authors' published code and settings on the same blockspace-filtered USPTO-MIT train and validation splits. Each archive contains the checkpoint, the upstream license and a note describing the training and the changes made to the authors' code. These checkpoints are distributed under the licenses of the respective upstream projects (MIT, MIT and BSD-3-Clause, Apache-2.0 and MIT). All other files are licensed under the MIT License (see LICENSE). The USPTO-MIT splits are by Jin et al. (2017) and were released under the MIT License, whose copyright notice is included in the data archive.