Finding Support Examples for In-Context Learning (Findings of EMNLP 2023)

The official repository for paper "Finding Support Examples for In-Context Learning" (EMNLP 2023 Findings). Check out our paper for more information.

Datasets

Download the train_data and test_data from Google Drive, and put them in the current folder.

Installation

conda create -n icl_support_examples python=3.8
conda activate icl_support_examples
conda install pytorch=1.7.1 -c pytorch
pip install transformers==4.3.0
pip install fitlog

Run Our Method

# dataset=[sst2,sst5,amazon_b,mr,subj,agnews,trec,dbpedia]
CUDA_VISIBLE_DEVICES=0 bash commands/run_[dataset].sh

and then see the performance through fitlog as:

fitlog log fitlog_search_examples

Then see fitlog in your browser.

Citation

@article{icl_support_example,
  author       = {Xiaonan Li and
                  Xipeng Qiu},
  title        = {Finding Supporting Examples for In-Context Learning},
  journal      = {CoRR},
  volume       = {abs/2302.13539},
  year         = {2023},
  url          = {https://doi.org/10.48550/arXiv.2302.13539},
  doi          = {10.48550/ARXIV.2302.13539},
  eprinttype    = {arXiv},
  eprint       = {2302.13539},
  timestamp    = {Tue, 28 Feb 2023 14:02:05 +0100},
  biburl       = {https://dblp.org/rec/journals/corr/abs-2302-13539.bib},
  bibsource    = {dblp computer science bibliography, https://dblp.org}
}