Environments for Agent Lightning

1. Overview

This directory contains several environments and training datasets for reinforcement learning (RL) experiments.

Supported Environments

• ALFWorld
  An embodied instruction-following environment built on top of ALFRED, focusing on language-guided household tasks.
  https://github.com/alfworld/alfworld

• ScienceWorld
  A text-based scientific reasoning environment designed for training and evaluating agents on procedural and commonsense science tasks.
  https://github.com/allenai/ScienceWorld

---

 

2. Setup

Set up the required environments and dependencies for each environment using the provided scripts.

ALFWorld

# The setup script performs the following:
# - Creates a new conda environment named 'alfworld'
# - Installs all required ALFWorld dependencies

bash setup/setup_alfworld.sh

ScienceWorld

# The setup script performs the following:
# - Creates a new conda environment named 'sciworld'
# - Installs all required ScienceWorld dependencies

bash setup/setup_sciworld.sh

---

 

3. Preparing the Dataset

You can either use the provided datasets or create custom datasets by modifying the dataset generation scripts.
All datasets are located under the task_data/ directory.

ALFWorld

Using the Provided Dataset

• Training set: task_data/alfworld/train.parquet
• Test set: task_data/alfworld/test.parquet

Creating a Custom Dataset

• Modify the following script: task_data/alfworld/make_alfworld_dataset.py

---

ScienceWorld

We provide two types of datasets for ScienceWorld: single-task and multi-task. For more information about the types of ScienceWorld tasks and their contents, refer to task_data/scienceworld/scienceworld_tasks.json, which is generated by running task_data/scienceworld/collect_tasks.py.

1. Single-Task Dataset

Each task has its own dataset. To identify which task corresponds to a given task_num, please refer to task_data/scienceworld/split_Sets/id2taskname.json.

Using the Provided Dataset

• Training set: task_data/scienceworld/single_data/{task_num}/train.parquet
• Test set: task_data/scienceworld/single_data/{task_num}/test.parquet

Creating a Custom Dataset

• Modify the following script: task_data/scienceworld/make_single_sciworld_dataset.py

 

2. Multi-Task Dataset

This dataset combines multiple ScienceWorld tasks into a single dataset.

Using the Provided Dataset

• Training set: task_data/scienceworld/multi_data/train.parquet
• Test set: task_data/scienceworld/multi_data/test.parquet

Creating a Custom Dataset

• Modify the following script: task_data/scienceworld/make_multi_sciworld_dataset.py