This section contains demonstrations of Adaptor's applications for various use-cases.
- adapted_named_entity_recognition constructs a standard single-objective token classification training pipeline and compares it to the common multi-objective approach applying unsupervised adaptation (Masked Language Modeling objective) preceding the final objective (Token Classification objective). It also shows how to ease evaluation of the persisted model using objectives.
- new_objective_development shows you a guideline to develop a new training Objective for NMT in the Adaptor environment.
- robust_classification trains the NLI (Natural Language Inference) classifier on MNLI and evaluates the model on a held-out adversarial objective (HANS) during the training. We assess whether the model selection based on in-domain accuracy (MNLI) also brings the most-robust model and what in-domain price we need to pay to make the model robust to other dataset (HANS).
- simple_sequence_classification compares classification and generation approaches in single-objective trainings. A go-to, if you want to use the comfort of Adaptor interface for a simple use-case. On the way, it shows how easy it is in Adaptor to exchange one objective for another.
- unsupervised_machine_translation uses Adaptor to assess the gap between the supervised and a commonly-used unsupervised objective (BackTranslation).