This is the official implementation of the paper "Effective In-Context Example Selection through Data Compression" based on PyTorch.

Check the following instructions for reproducing experiments.

You can get the results of BM25 or sentence-bert.

python stage1_bm25.py

python stage1_sentbert.py

Note: You need to interrupt the forward propagation of GPT2 at the first layer firstly by changing the modeling_gpt2.py, which in the pre-downloaded file.

You can get the examples of different datasets and models by the following instructions.

For example:

python stage2_ours.py --data_type cola --model gpt2
python stage2_ours.py --data_type sick --model gpt2-medium

We referred to https://github.com/juny116/ICL-DeepSpeed testing process for testing.

You can modify the config file to test different tasks. For example, you can test with the following command:

cd ICL-DeepSpeed-main
python single_main_ours.py

We conducted the experiments based on the following environments:

• CUDA Version: 11.4
• torch version: 1.10.0
• OS: Ubuntu 18.04.5 LTS
• GPU: NVIDIA Geforce RTX 3090
• CPU: Intel(R) Xeon(R) Silver 4214 CPU @ 2.20GHz