Paper: Explore Spurious Correlations at the Concept Level in Language Models for Text Classification (ACL 2024)
Link: https://arxiv.org/pdf/2311.08648
For distilbert model: python train_bert.py --dataset amazon-shoe-reviews --concept size --method biased
For llama2 model: python llama2_classification.py --lora_r 8 --epochs 3 --dropout 0.1 --pretrained_ckpt [llama2_checkpoint_path] --dataset amazon-shoe-reviews --concept size --method biased
dataset name: amazon-shoe-reviews, imdb, yelp_polarity, cebab, boolq
method name: original, biased, downsample, upsample, mask