Murphyyyy/RIGen-Xonotic-Dataset

RIGen Xonotic Dataset

RIGen is a large, manifest-driven multimodal gameplay dataset for learning and evaluating interactive game-world dynamics. It is collected from the Xonotic first-person shooter and pairs rendered audiovisual observations with player actions, opponent motion, combat outcomes, visibility, prompts, and structured game state.

This repository is the RIGen Paper Open Release v1. It contains the final files referenced by the four paper data components. The package is self-contained at the file level: manifest paths are relative to the repository root and the files are ordinary independent copies.

Dataset contents

Component	Manifest records	Files	Size (decimal)	Main purpose
Basic Dynamics	50,140	250,700	365.39 GB	Local player-controlled audiovisual dynamics and state/action supervision
Interaction short	42,000	256,201	120.77 GB	Short opponent-motion interaction blocks with actions, outcomes, visibility, and prompts
Interaction-Long	1,791	7,401	77.40 GB	Longer multi-block interaction trajectories
Arena	1,800	7,207	142.18 GB	Duel/arena trajectories with combat events, state transitions, and prompt/rollout metadata
Total	95,731	521,509	705.74 GB

The manifest record counts describe rows in the component manifests; they are not necessarily episode counts. The exact file and row inventories are recorded in release_summary.json.

Repository layout

README.md
LICENSE.md
NOTICE
UPSTREAM_LICENSES/
manifests/                 # component JSONL manifests
provenance/               # source-to-package provenance maps
files/basic_dynamics/     # sharded media, tables, and metadata
files/interaction_short/
files/interaction_long/
files/arena/
release_summary.json

Large file collections are split into numeric shard directories so that no single directory exceeds the Hugging Face repository directory limit. The shard name is an implementation detail; use manifest paths instead of reconstructing paths from episode IDs.

File types and supervision

The release contains combinations of rendered video/audio media, frame and action tables, scenario/prompt JSON, QA records, and provenance metadata. Depending on the component, records expose fields such as:

• observations and rendered clips;
• continuous movement/view controls and button actions;
• player and opponent motion/state;
• combat, hit, death, and termination events;
• visibility and camera-conditioning labels;
• natural-language prompts and environment/appearance descriptions;
• split, acceptance, QA, and source provenance fields.

The JSONL manifests are the authoritative join layer. Each path in a manifest points to a file under files/ using a repository-relative path.

How to use

Clone or download the repository with Git LFS support, then read a component manifest as JSONL. For example:

import json
from pathlib import Path

root = Path("/path/to/RIGen-Xonotic-Dataset")
with (root / "manifests/interaction_short.jsonl").open() as f:
    record = json.loads(next(f))

video = (root / "manifests" / record["video_path"]).resolve()
action = (root / "manifests" / record["action_path"]).resolve()
print(video, action)

The paths are resolved relative to the manifests/ directory. For reproducible training, preserve the manifest row, split, and provenance fields together with any derived preprocessing.

This is a file-oriented release rather than a single Arrow table, so it is intended to be consumed through the provided manifests. The package intentionally excludes raw capture directories, .dem replays, source_64hz.mkv, logs, and production intermediates.

Reproducibility and code

The matching minimal construction code release contains the scripts, configurations, patches, and modified QuakeC source needed to recreate these four components. The data package itself does not include the development archive or engine build tree. Keep the release version and manifest hashes fixed when reproducing experiments.

Licensing and attribution

This is a component-licensed release. RIGen-authored annotations, action/state/event/visibility labels, prompts, manifests, QA metadata, and documentation are released under CC BY 4.0. Xonotic-derived gameplay media and reference assets remain subject to the applicable upstream Xonotic and asset-level terms; they are not relicensed as CC BY 4.0.

Read LICENSE.md before redistribution and retain NOTICE and UPSTREAM_LICENSES/. The upstream Xonotic project is available at github.com/xonotic/xonotic.

Citation

If you use this dataset, please cite the RIGen paper and identify the dataset release as RIGen Paper Open Release v1. The paper citation will be added here when the public paper record is finalized.

Known limitations

The media is derived from a game environment and is subject to the upstream game's content and licensing terms. The four components use different sampling and supervision contracts; do not merge their manifests without preserving the component name and schema fields. Some labels are generated from authoritative game telemetry and some prompts are derived from accepted render/state records, so a prompt should not be treated as an independent ground-truth observation.