micm_nlp.training¶
Training: the run driver, Trainer subclasses, callbacks, collators and batching.
runner—TRAINER: builds the HuggingFaceTrainerfrom the config — arguments, collator, callbacks, evaluation — and runs it.trainers—CustomTrainerMixinandRandomTaskExclusionBatchSampler;custom_trainer_class_factorymixes the former into whicheverTrainerclasstrainer.clsnames.callbacks—CustomEarlyStoppingCallback,ParamNormLogger,NormalizePromptEncoderEmbeddings,DownstreamFineTuningCallback,EmptyCudaCacheCallback.batching—TokenBudgetBatchSamplerandcalibrate_token_budget, for batching by token count rather than by row count.data_collators— collators for PLM, seq2seq-with-shifted-labels, and task-id-decorated batches.logits_processors—ConstrainedPrefixLogitsProcessor, applied at generation time.