This document describes a dataset of human NetHack gameplay curated from the publically available and permissively licensed recordings available at http://nethack.alt.org.
The original data consists of a very large collection of human gameplay in the form of raw TTY terminal command recordings, which is a complex format to ingest. We release a processed version of a subset of the dataset, with the following transformations:
We have used python-based terminal emulator pyte (https://github.com/selectel/pyte) to decode the serialized terminal commands into fixed-size tensors of terminal characters, colors, backgrounds, and cursor locations. We remove any sessions that use non-default user interface settings, including terminal sizes larger than the standard 24x80, interface configurations such as Curses, and non-ASCII character remappings. This brings the data format closer to the observations that would be seen by an agent using the NetHack Learning Environment (NLE: https://github.com/facebookresearch/nle). We have curated the dataset (which consists of over 7 million gameplay sessions, taking hundreds of terabytes of storage space once decoded) by filtering the dataset to include only the gameplay from the top 10 NetHack players as ranked by Z-score (https://nethackwiki.com/wiki/Z-score), a measure of player ability. We annotate each gameplay session with the following metadata where it has been possible to parse from the observations:
| Metadata | Description |
|---|---|
alignment |
Character alignment |
length |
Number of timesteps in the session |
max_dungeon_number |
Maximum dungeon number reached |
max_exp_level |
Maximum experience level reached |
max_exp_points |
Maximum experience points achieved |
max_level_number |
Maximum dungeon level reached |
min_dungeon_number |
Minimum dungeon number reached |
min_exp_level |
Minimum experience level reached |
min_exp_points |
Minimum experience points achieved |
min_level_number |
Minimum dungeon level reached |
rank |
First character rank achieved |
role |
Character role (class) |
stamp |
Unix timestamp when the session was started |
username |
Username of the player that recorded the session |
The metadata is available as a pickle archived dictionary that maps session IDs to metadata dictionaries with the keys described above.
The full curated subset contains 16478 gameplay sessions. Note that a gameplay session starts when the user logs in, and ends when the user logs out. A complete episode can therefore consist of multiple consecutive sessions. The dataset in total amounts to approximately 184M timesteps, and is approximately 12GB in size.
The entire dataset can be downloaded as a single file here: nao_top10.tar [11.8 GB]
We also provide separate access to the metadata and recorded gameplay of
individual users to allow for downloading smaller subsets of the dataset (note
that all of the following files are included in nao_top10.tar above):
| File | Size |
|---|---|
| metadata.pkl | 10 KB |
| 78291.tar | 10.1 MB |
| Fek.tar | 1.1 GB |
| Luxidream.tar | 8.2 MB |
| Stroller.tar | 3.8 GB |
| Tariru.tar | 1.5 GB |
| YumYum.tar | 1.8 GB |
| oh6.tar | 1.7 MB |
| rschaff.tar | 29.3 MB |
| stenno.tar | 1.3 MB |
| stth.tar | 3.6 GB |
All of the data originates from http://nethack.alt.org and is licensed under a Creative Commons Attribution 4.0 International License.
The data is organized as [username]/[session_id].npz where session_id is a
base64 string, and the file is in .npz format, which can be loaded using
numpy:
session = np.load(filename) # Dict of observations
tty_chars = session['tty_chars'] # Shape = (T, 24, 80)
tty_colors = session['tty_colors'] # Shape = (T, 24, 80)
tty_background = session['tty_background'] # Shape = (T, 24, 80)
tty_cursor = session['tty_cursor'] # Shape = (T, 2,)
The tty_chars, tty_colors, and tty_cursor have similar semantics to the corresponding observations from the NLE environment. tty_background corresponds to the background color used to represent the specials observation in NLE when rendered in the browser, although since specials is a bitfield that can have multiple bits enabled, the background colors do not map directly to specials.
We also include a short rendering utility that uses PyGame to generate images of arbitrary states, for ease of visualizing the data. The code will generate images of states from the dataset, as well as observations from the NLE environment. The renderer can also take arbitrary foreground and background color arrays, which we use in the paper to generate saliency visualizations and to highlight differences between states. The renderer can be used to visualize observations from the dataset as follows:
image_renderer = ImageRenderer()
session = np.load(filename)
# Get the fifth observation of the session as a dict.
# We choose the fifth frame here because the first few are just setup.
frame_num = 5
observation = dict(
tty_chars=session['tty_chars'][frame_num],
tty_colors=session['tty_colors'][frame_num],
tty_background=session['tty_background'][frame_num],
tty_cursor=session['tty_cursor'][frame_num])
img = image_renderer.render(observation)
Install cmake:
sudo apt install cmake
Install nle, numpy, and pygame:
pip install -r ./requirements.txt
If you found this dataset useful for academic work, then please cite the corresponding paper: Learning About Progress From Experts.
@inproceedings{brucelearning,
title={{Learning About Progress From Experts}},
author={Bruce, Jake and Anand, Ankit and Mazoure, Bogdan and Fergus, Rob},
booktitle={International Conference on Learning Representations},
year={2023}
}