Infatoshi/terraria-gameplay-256x144-15fps

Terraria Gameplay Corpus (256x144, 15 fps)

841 hours of vanilla Terraria gameplay video: 310 videos from 141 YouTube channels, filtered for clean side-view gameplay and re-encoded to a compact, uniform format for video pretraining (VPT-style behavioral priors, world models, video prediction).

• Format: H.264 MP4, 256x144, 15 fps, no audio (-cq 23, yuv420p)
• Size: ~33 GB total
• Videos: 310, named videos/<youtube_id>.mp4
• Duration: 30 min minimum per video, 2.7 h average, 25.1 h max
• Content: vanilla Terraria playthroughs and longplays (modded content filtered out)

Why this exists

Terraria is a side-view sandbox platformer with gravity, mining, building, and ranged combat. Existing RL benchmarks for sandbox games (Crafter, Craftax, MineRL) are either top-down or a different game entirely. This corpus provides large-scale human gameplay footage in a uniform low-resolution format suitable as pretraining data for agents that play the real game.

Files

videos/<id>.mp4            310 processed videos, named by YouTube video ID
metadata.tsv               per-video: id, source URL, channel, title, source resolution, duration (s)
provenance/
  pass1_decisions.tsv      metadata-filter decisions for all downloaded videos
  pass2_decisions.tsv      SigLIP per-video frame-class fractions + keep/drop
  vlm_decisions.tsv        Qwen2.5-VL game-identity votes + keep/drop

Collection and cleaning pipeline

1. Download: candidate Terraria playthrough/longplay videos fetched with yt-dlp at <=480p (most sources are 854x480 or 640x360).
2. Pass 1 - metadata filter (ffprobe + title rules): dropped videos shorter than 30 minutes, below 480x270, with aspect ratio outside 1.5-2.0, or with modded-content titles (blacklist of 18 popular mods: Calamity, Thorium, tModLoader, etc.). 320 videos survived (905 h).
3. Pass 2 - frame classifier (SigLIP base, 12 uniformly spaced frames per video, zero-shot against 6 text prompts: gameplay / facecam / intro / menu / stream overlay / blank): dropped videos with >25% facecam frames, >50% intro+menu+blank frames, >40% overlay frames, or <40% gameplay frames.
4. Pass 3 - game identity (Qwen2.5-VL-7B, 3 frames per video, "Is this Terraria? YES/NO" majority vote): dropped videos identified as other games or with mixed/uncertain votes.
5. Re-encode: survivors re-encoded with ffmpeg/NVENC to 256x144 @ 15 fps, H.264 -cq 23, audio stripped. Final count: 310 videos, 841 hours.

All filter decisions, thresholds, and per-video classifier scores ship in provenance/ so the filtering is auditable and reproducible.

Known limitations

• Filtering thresholds are permissive: kept videos may still contain brief intros, menus, map/inventory screens, small corner facecams, or stream overlays below the drop thresholds.
• The modded-content filter relies on titles plus VLM frame checks; occasional lightly modded or texture-packed footage may remain.
• 256x144 is intentionally low resolution; small UI text is not legible.
• No action labels - this is video-only data. Pair with an inverse dynamics model for VPT-style training.

Usage

from huggingface_hub import snapshot_download
path = snapshot_download("Infatoshi/terraria-gameplay-256x144-15fps", repo_type="dataset")

Then decode frames with your tool of choice (decord, PyAV, torchvision):

import csv
with open(f"{path}/metadata.tsv") as f:
    meta = list(csv.DictReader(f, delimiter="\t"))
print(meta[0]["title"], meta[0]["duration_s"])

Attribution and takedown

The gameplay footage was created by 141 YouTube creators; metadata.tsv links every video to its source URL and channel. Copyright in the underlying videos remains with the original creators. Terraria is a trademark of Re-Logic; this dataset is not affiliated with or endorsed by Re-Logic.

This dataset is shared for research purposes (agent pretraining, world models, video prediction). If you are a creator whose video is included and you want it removed, open a discussion on this repo and it will be taken down promptly.