Invariant Language Modeling

Implementation of the training for invariant language models.

Motivation

Modern pretrained language models are critical components of NLP pipelines. Yet, they suffer from spurious correlations, poor out-of-domain generalization, and biases. Inspired by recent progress in causal machine learning, we propose invariant language modeling, a framework to learn invariant representations that should generalize across training environments. In particular, we adapt IRM-games to language models, where the invariance emerges from a specific training schedule in which environments compete to optimize their environment-specific loss by updating subsets of the model in a round-robin fashion.

Model Description

The data is assumed to come as n distinct environments and we aim to learn a language model that focusing on correlations that generalize across environments.

The model is decomposed into two components:

• ϕ the main body of the transformer language model,
• w the language modeling head that predicts the missing token.

In our implementation, there are now as many heads as environments: n. For each data point, all heads make their predictions and they are averaged. However, during training we sample one batch from each environment in a round-robin fashion. When seeing a batch from environment e only the head w_e and the main body ϕ receive a batch update.

Usage

To get started with the code:

pip install -r requirements.txt

PyTorch with a CUDA installation is required to run this framework. Please find all useful installation information here

Then, to continue the training of a language model from a huggingface checkpoint:

python3 run_invariant_mlm.py \
    --model_name_or_path roberta-base \
    --validation_file data-folder/validation_file.txt \
    --do_train \
    --do_eval \
    --nb_steps 5000 \
    --learning_rate 1e-5 \
    --output_dir folder-to-save-model \
    --seed 123 \
    --train_file data-folder/training-environments \
    --overwrite_cache

Currently, the supported base models are:

• roberta: checkpoints
• distilbert: checkpoints

Implementation

To train language models according to the IRM-games, one needs to modify:

• the training schedule to perform batch updates according to each environment in a round-robin fashion. This logic is implemented by the InvariantTrainer in invariant_trainer.py', a class inherited from the Trainer` from huggingface.
• the language modeling heads in the model. It needs one head per environment. This is done by creating variations of the base model classes. It is implemented in invariant_roberta.py for roberta and in invariant_distilbert.py for distilbert.

Contact

Maxime Peyrard, [email protected]