LaMoTO: Language Modelling Tasks as Objects

Language Modelling Tasks as Objects (LaMoTO) provides a framework for language model training (masked and causal, pretraining and finetuning) where the tasks, not just the models, are classes themselves. It abstracts over the HuggingFace transformers.Trainer with one goal: reduce the entire model training process to a single method call task.train(hyperparameters).

pip install lamoto[all]

Note: to be able to use the W&B integration, make sure you first run wandb login in a command-line terminal on the system you want to run on.

Usage

Let's say you want to train a RoBERTa-base model for dependency parsing (for which, by the way, there is no HuggingFace class). This is how you would do that in LaMoTO, supported by the magic of ArchIt:

from archit.instantiation.basemodels import RobertaBaseModel
from archit.instantiation.heads import DependencyParsingHeadConfig, BaseModelExtendedConfig
from lamoto.tasks import DP
from lamoto.training.auxiliary.hyperparameters import getDefaultHyperparameters

# Define task hyperparameters.
hp = getDefaultHyperparameters()
hp.model_config_or_checkpoint = "roberta-base"
hp.archit_basemodel_class = RobertaBaseModel
hp.archit_head_config = DependencyParsingHeadConfig(
    head_dropout=0.33,
    extended_model_config=BaseModelExtendedConfig(
        layer_pooling=1
    )
)

# Instantiate language modelling task as object, and train model.
task = DP()
task.train(hyperparameters=hp)

Features

• Train models on >15 pre-training/fine-tuning tasks. See a list by importing from lamoto.tasks.
  • Model architectures come from ArchIt, which means that as long as you have a BaseModel wrapper for your language model backbone, you can train it on any task, regardless of whether you wrote code defining the backbone-with-head architecture required for that task.
  • Custom (i.e. given) architectures are also supported.
• Evaluate models with a superset of the metrics in HuggingFace's evaluate, with custom inference procedures (see e.g. strided pseudo-perplexity or bits-per-character).
• Augment datasets before training or evaluating by somehow perturbing them.
• Supports TkTkT tokenisers.
• W&B integration.

Alternative packages

There exist other libraries that abstract across training tasks in an effort to avoid heavily dedicated training scripts. I'm aware of the following packages (although I'm not sure how extensible they are):

• jiant
• itrain
• machamp