Tuning Free OOD Detection

This repository contains the official code for the paper "Is Fine-tuning Needed? Pre-trained Language Models Are Near Perfect for Out-of-Domain Detection". (TODO: Add link once on arxiv)

---

Setup

First, create a virtual environment for the project (we use Conda to create a python 3.9 environment) and install all the requirments using requirements.txt.

1. conda create -n ood_det python==3.9
2. conda activate ood_det
3. pip install -r requirements.txt

Now, create two directories:

1. Data directory
2. Model directory - create two subdirectories: pretrained_models and finetuned_models

Both directories can be located anywhere but should be specified in config.py (See more below).

---

Running Experiments

At the start of each session, run the following: . bin/start.sh

Running OOD Detection on a Model

Directory paths and training arguments can be specified in config.py. Some important arguments are:

• DATA_DIR: Path to the data directory
• MODEL_DIR: Path to the model directory
• task_name: Name of the dataset to use as in-distribution (ID) data.
• ood_datasets: List of datasets to use as out-of-distribution (OOD) data.
• model_class: Type of model to use. Options are roberta, gpt2and t5 (base versions of all models used).
• do_train: If true, trains the model on the ID data before performing OOD detection.

After specifying the arguments, run the following command: python run_ood_detection.py

Training a Model through TAPT

Running the command below will extend the pretraining process on a specified dataset. Once again, the values in config.json can be used to specify the dataset and model to be used. python tapt_training/pretrain_roberta.py

After the new model has been saved to the pretrained_models directory within the MODELS_DIR directory (specified in config.py), it can be used for OOD detection by running run_ood_detection.py.

---

Citation

If you find this repo helpful, you are welcome to cite our work:

 @inproceedings{uppaal2023fine,
      title={Is Fine-tuning Needed? Pre-trained Language Models Are Near Perfect for Out-of-Domain Detection}, 
      author={Rheeya Uppaal and Junjie Hu and Yixuan Li },
      booktitle = {Annual Meeting of the Association for Computational Linguistics},
      year = {2023}
}

Our codebase borrows from the following:

@inproceedings{zhou2021contrastive,
  title={Contrastive Out-of-Distribution Detection for Pretrained Transformers},
  author={Zhou, Wenxuan and Liu, Fangyu and Chen, Muhao},
  booktitle={Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing},
  pages={1100--1111},
  year={2021}
}

@article{liu2020tfew,
  title={Few-Shot Parameter-Efficient Fine-Tuning is Better and Cheaper than In-Context Learning},
  author={Liu, Haokun and Tam, Derek and Muqeeth, Mohammed and Mohta, Jay and Huang, Tenghao and Bansal, Mohit and Raffel, Colin},
  journal={arXiv preprint arXiv:2205.05638},
  year={2022}
}