This repository contains the codes for the EMNLP 2023 Findings paper, FinePrompt. Our work deals with the idea of transferring task-specific inductive biases from finetuned models to prompts, as a way of improving GPT-4’s compositional reasoning capabilities.

Since this work only deals with the effective prompting on GPT-4, our work requires the installation of the openai API.

pip install openai

python main.py --bias_type <bias_type_name> --dataset <dataset_name>

where --bias_type can be one of [genbert, numnet, decomprc, quark, sae, qdgat] and --dataset can be one of [drop, musique]

The set of base and in-context prompts are provided in the prompts.py.

@inproceedings{kim2023fineprompt,
        author={Jeonghwan Kim and Giwon Hong and Sung-Hyon Myaeng and Joyce Jiyoung Whang},
        title={FinePrompt: Unveiling the Role of Finetuned Inductive Bias on Compositional Reasoning in {GPT}-4},
        booktitle={Findings of the Association for Computational Linguistics: EMNLP 2023},
        year={2023},
        pages={3763--3775}
}