Errata
In this section, we will document changes that affected a method’s correctness.
Although a lot of effort is spent in this repository to ensure correctness (through extensive unit and integration testing), errors will happen. The best we can do is to leverage the shared knowledge of all contributors, to spot and fix these issues. Most importantly, we provide full transparency by documenting such cases here.
In case you are using small-text in a scientific context, make sure to document the version that has been used.
Classifiers
TransformerBasedClassification: parameter groups were omitted when using the layer-specific fine-tuning functionality (#38; fixed in v1.3.1).
SetFitClassification: certain training arguments (such as num_epochs and max_steps) were not passed correctly (#79; affects small-text v2.0.0dev1, v2.0.0dev2 with setfit>=1.0.0; fixed in v2.0.0.dev3 / v2.0.0).
Query Strategies
BADGE: the initial implementation was incorrect (fixed in v2.0.0.dev4).
