Awesome

Contrastive Novelty-Augmented Learning

This is the official code for the paper CoNAL: Anticipating Outliers with Large Language Models. This repository contains the code for replicating the all experiments including data splitting, generation, and augmented training, and is extensible with other OOD detection/selective prediction methods. Broadly, we generate OOD sets of examples from novel classes, then augment our training set and train with a confidence contrastive loss to reduce novel class overconfidence.

Dependencies

We train classifier models using HuggingFace Transformers. For label generation, we use the OpenAI GPT-3 API. You will need to register for an account on the website to use GPT-3. Fortunately we do not need to make a large number of API calls, so the total cost of running the method is not high.

Install

1. Clone the repo using git clone https://github.com/albertkx/CoNAL.git
2. Make a new virtual environment with Python 3.8.
3. Run pip install -r requirements.txt.
4. Download TACRED following https://nlp.stanford.edu/projects/tacred/#access.
5. Add your OpenAI API key in api_keys.py
6. You may need to set up your Weights&Biases account for logging, following Step 1 here

Folder Structure

• preprocessing/ contains code for creating training splits for the datasets we evaluate on: TREC-10, Emotion, AGNews, and TACRED, and additionally downloads Wikitext.
• ood_generation/ contains code for generating OOD sets using novelty prompting or few-shot prompting on a LLM generator.
• training/ contains code for training a classifier model augmented with an OOD set.
• scoring/ contains utilities for scoring many models at once.

Additional folders will be created for the following:

• All training data splits will be located in data/.
• All generations will be saved to generations/.
• All model checkpoints will be saved to models/.

Running Experiments

Examples of running all experiments are shown in the Makefile. Below are more specific details.

Replicating Data Splits

To create both create the train data splits and download Wikitext:

python3 -m preprocessing.make_data
	--out_path data/
	--tacred_data_path /path/to/tacred/

Generating OOD Sets

We provide three interchangeable scripts for generating OOD data, generate_np.py, generate_zeroshot.py, and generate_fewshot.py. All of them can be called with roughly the same parameters. Below is an example of generating 1000 OOD examples on the split 0 (where the class 0 is held-out) of the Emotion dataset. The names given to the model sizes can be found in generation_utils.py.

CUDA_VISIBLE_DEVICES=0 srun --gres=gpu:1 -t 72:00:00 python3 -m ood_generation.generate_np \
	--dataset emotion \
	--gold_ood_class 0 \
	--model_size large \
	--num_generations 1000 \
	--generation_batch_size 16 \
	--example_generator_size large \
        --label_generator_size large \
	--output_dir_name test \
	--do_moby \
	--seed 1

By default, label and example generation use GPT-J 6B. Both of these generator models can be easily swapped out by changing parameters, e.g., by passing gpt3 instead of large.

Training Models

After generating an OOD example set, we can use this to augment the existing train set. To train with confidence contrastive loss:

CUDA_VISIBLE_DEVICES=0 python3 -m training.train_classifier
	data/trec10/0/
	models/trec10/0/bert_ccl_seed_1/
	--ood_training_type ccl
	--ood_path data/wikitext/
	--model_type bert_base
	--num_training_steps 5000
	--checkpoint_steps 100
	--seed 1
	--do_loss_log

We also implement Outlier Exposure training which can be called roughly the same way. See the recipe in the Makefile for how to run this: make train_oe.

Scoring

To score all models of name model_name_here on all splits of dataset dataset_name, run CUDA_VISIBLE_DEVICES=0 python3 -m scoring.score model_name_here dataset_name. If your folder structure is different, you may need to modify scoring/score.py.

Citation

@inproceedings{xu2023conal,
  author    = {Albert Xu and Xiang Ren and Robin Jia},
  title     = {CoNAL: Anticipating Outliers with Large Language Models},
  booktitle = {arXiv},
  year      = {2022},
}