jjhw / rot

Official code for ACL 2023 (short, findings) paper "Recursion of Thought: A Divide and Conquer Approach to Multi-Context Reasoning with Language Models"

Recursion of Thought

Official PyTorch implementation of ACL 2023 (short, findings) paper: Recursion of Thought: A Divide and Conquer Approach to Multi-Context Reasoning with Language Models.

• Paper
• Poster

Installation

We recommend using Anaconda. The following command will create a new conda environment rot with all the dependencies.

conda env create -f environment.yml

To activate the environment:

conda activate rot

Directory structure

├── configs
├── data
├── episodes
├── models
├── eval.py
├── gpt_eval.py
├── gpt_fine_tune.py
├── train_cmds
└── train.py

• configs: Model & training configurations
• data: Codes related to problem generation
• episodes: Problem configurations
• models: Codes for Transformer and LSTM
• eval.py: Evaluation logic
• gpt_eval.py: GPT-3 evaluation script
• gpt_fine_tune.py: GPT-3 fine-tuning script
• train_cmds: The commands used to train the models
• train.py: The main training logic

GPT-3 experiments

Run the gpt_fine_tune.py script to create data for GPT-3 and request fine-tuning. The API key should be present in the environment variable OPENAI_API_KEY.

python gpt_fine_tune.py

The following script will evaluate the fine-tuned models.

python gpt_eval.py

Experiments with the tiny models

Use the following command to train a model.

python train.py -c [config_path] -e [episode_path] -p [paradigm] -l [log_directory]

config_path should be one of the files in configs directory, and episode_path should be one of the files in episodes. paradigm should be one of wt, cot, and rot. In the train_cmds file, we provide all commands that we used to for our main experiments. The evaluation results will be recorded in the TensorBoard summary.