IRT-Router

View on GitHub

Experiments

Training

We provide training data in the following file:

  • data/train.csv: Training dataset.

You can train the M-IRT router using the following command:

python train_mirt.py

Similarly, to train the N-IRT router, run:

python train_nirt.py

We also provide a trained model checkpoint:

  • mirt_bert.snapshot: uses bert-base-uncased as embedding model.

Testing

  • data/test1.csv: In-distribution test set.
  • data/test2.csv: Out-of-distribution test set.

To evaluate the M-IRT router on the in-distribution test set, use the following command:

python test_router.py --router mirt --emb_name bert --test_path test1 --a 0.8 --lamda 0.3

Alternatively, you can execute the pre-written script:

sh test.sh

Router Usage

To be continued…