Language Models are Few-Shot Butlers

This repository contains the code to reproduce the experiments in Language Models are Few-Shot Butlers.

Installation

ALFWorld

$ git clone https://github.com/alfworld/alfworld.git
$ cd alfworld
$ git checkout c0501c807fa77dd25ecfc86e92cd782226f5c74f
$ sed -i '/torch/d' requirements.txt 
$ pip3 install -r requirements.txt
$ python3 setup.py develop

PyTorch

$ pip3 install torch==1.8.1+cu111 torchvision==0.9.1+cu111 -f https://download.pytorch.org/whl/torch_stable.html

Specific dependencies

$ pip3 install -r requirements_butlers.txt

Action modeling

Training a GPT-2 medium (345M parameters) model with action modeling on expert demonstrations:

$ python3 src/action_modeling.py configs/action_modeling_config.yaml

Evaluation

Evaluating the resulting model on out-of-distribution task instances:

$ python3 src/alfworld_trainer.py configs/alfworld_trainer_eval_config.yaml

Models

Trained models are hosted on the Hugging Face hub. For instance, set model_path to vmicheli/lm-butlers-gpt in alfworld_trainer_eval_config.yaml to skip the Action modeling step.