MoT: Memory-of-Thought Enables ChatGPT to Self-Improve (EMNLP 2023)

The official repository for MoT. Check out our paper for more information.

Requirements

Python>=3.8

pip install torch==1.8.2+cu111 torchtext==0.9.2 -f https://download.pytorch.org/whl/lts/1.8/torch_lts.html
pip install -r requirements.txt

Datasets

Download the datasets from Google Drive.

Run Our Method

First, fill your account in to a text file, as follows:

[Email_1]----[Password_1]----[Openai_API_Key_1]
[Email_2]----[Password_2]----[Openai_API_Key_2]
...

Our code supports using multiple openai accounts simultaneously and call openai api in parallel.

If you want to run the entire method:

# dataset:[aqua, drop, anli_a1, anli_a2, anli_a3, obqa, com_v, boolq, fact_checker, qa_wikidata]
dataset=[dataset] bash commands/run_mot_full.sh

Or you can directly download the memory of thoughts in pre-thinking, from Google Drive, and run the subsequent memory filtering and recalling:

# dataset:[aqua, drop, anli_a1, anli_a2, anli_a3, obqa, com_v, boolq, fact_checker, qa_wikidata]
dataset=[dataset] bash commands/run_mot_with_existing_memory.sh

Citing MoT

@article{memory_of_thought,
  author       = {Xiaonan Li and
                  Xipeng Qiu},
  title        = {MoT: Memory-of-Thought Enables ChatGPT to Self-Improve},
  journal      = {CoRR},
  volume       = {abs/2305.05181},
  year         = {2023},
  url          = {https://doi.org/10.48550/arXiv.2305.05181},
  doi          = {10.48550/ARXIV.2305.05181},
  eprinttype    = {arXiv},
  eprint       = {2305.05181},
  timestamp    = {Fri, 12 May 2023 16:06:58 +0200},
  biburl       = {https://dblp.org/rec/journals/corr/abs-2305-05181.bib},
  bibsource    = {dblp computer science bibliography, https://dblp.org}
}