Code for Explanation Selection Using Unlabeled Data for Chain-of-Thought Prompting (EMNLP 2023).

Setup

• python==3.8
• requirements: pip install -r requirements.txt
• Set OPENAI KEY: export KEY=yourkey
• Run mkdir misc expl_search/misc expl_search/annotations expl_search/candidates

TLDR

We directly provide pairs of seed few-shot CoT prompts and searched CoT prompts in prompt pairs. (Note these prompts are optimized for code-davinci-002).

Run evaluation to compare seed CoTs and optimized CoTs on code-002
SETS="0" sh exp_scripts/exp_result.sh gsm # datasets include gsm, ecqa, esnli, strategyqa

Run Optimization Experiments

We take GSM as an example dataset

# SEED
sh exp_scripts/main/gsm.sh SEED
# OSACC
sh exp_scripts/main/gsm.sh OSACC
# OSLL
sh exp_scripts/main/gsm.sh OSLL

Citation

@InProceedings{Ye-Durrett:2023:explselect,
  title = {Explanation Selection Using Unlabeled Data for In-Context Learning},
  author = {Xi Ye and Greg Durrett},
  booktitle = {Proceedings of EMNLP},
  year = {2023},
}