zrustc / KNN-C

Official implementation of paper "Improving Few-Shot Performance of Language Models via Nearest Neighbor Calibration"

Official implementation of our paper "Improving Few-Shot Performance of Language Models via Nearest Neighbor Calibration"

Requirements

• python==3.6
• torch==1.10.1
• torch-scatter==2.0.7
• transformers==4.18.0

Datasets

We used the same datasets from Karimi Mahabadi et al. (ACL 2022). We first obtain datasets via pre-processing data scripts in https://github.com/facebookresearch/perfect, then follow cmd to construct datasets. Note that set the correct data-path before running scripts.

python3 tasks.py
sh data_process.sh

Training and Inference

• Download the pre-trained models (e.g., roberta-large)
• run sh task_run.sh