fix: update parameter name in workflow description

docs/guide/user-guide/workflow.md

@@ -327,7 +327,7 @@ The aim of our software is to produce a text containing as few errors as possibl
   -- Set the 'Number of folds to train' (i.e. the number of models to train) to 5. → Training will occur with a model package containing five individual models.
   -- **'Only train a single fold box':** please don't fill out this box!
   -- Set the **'Number of models to train in parallel'** at -1. → All training models will be trained simultaneously.
-  -- If all characters contained in the pretraining model need to be kept in the model you wish to train (i.e. added to its so called whitelist), please check the **'Keep all characters loaded from the last model'** box.
+  -- If all characters contained in the pretraining model need to be kept in the model you wish to train (i.e. added to its so called whitelist), please check the **'Keep codec of the loaded model(s)'** box.
   -- In effect, the **'Whitelist characters to keep in the model'** is the exhaustive list of characters used during training and in the subsequently generated model. Any character not contained in the whitelist won't be included in the process.
   -- **'Pretraining'**: Either **'Train each model based on different existing models'** (a menu will appear containing five dropdown lists. Inside each of them, enter one of the five models belonging to the model package used as advised earlier. Regardless of the training step (be it the first round or the third), always enter the five models used since the beginning) or **'Train all models based on one existing model'** (click on this setting if you started training using only one model. Simply select that exact training model for each repetition of the training process).
   -- **'Data augmentation':** Please don't fill out this box! This function describes the data augmentation per line. Users can enter a number, e.g. 5, in order to increase the amount of training material. This can lead to the generation of better performing models. However, this process is more time-costly than the standard route.