GG800-Subset 🎮✨

Welcome to GG800-Subset - a clean, action-conditioned GTA5 trajectory subset for world-model research.
Built with love at LucidML 💙

---

What is this?

GG800-Subset is a curated public slice of the larger GG800 pipeline.
Each sample is a synchronized triplet:

• capture.mkv -> the video clip
• events.srt -> human-readable control timeline overlay
• events.jsonl -> frame-level machine-readable control states

The key contract: frame and controls are aligned by frame index (not loose timestamp matching), so training pipelines can reliably map visuals <-> actions.

What the footage actually is 🎥

GG800-Subset contains action-agent gameplay footage from GTA V collected with a bias-reduction collection policy:

• capture environment uses GTA5 Enhanced with the NaturalVision Enhanced (NVE) visual mod for higher-fidelity graphics;
• each clip starts from a randomized world spawn location;
• the agent is reset into either a random pedestrian state or a random vehicle context (for example car/bike/other road vehicle);
• actions are programmatically generated to avoid repetitive route-locked behavior and improve behavioral coverage.

Each clip follows a fixed format:

• resolution: 720p
• frame rate: 12 FPS
• clip length: 384 frames (32 seconds)
• events.jsonl: exactly 384 lines, one per frame

So the alignment invariant is strict:
frame i in video <-> line i in events.jsonl for i = 0..383.

Practical note: the first 2 seconds (24 frames) may include scene/asset loading artifacts in some clips.
For model training/inference pipelines, a common practice is to skip these initial frames and use the final 360 frames.
Even when you apply this trim, action logs remain perfectly synchronized with frames by index.

---

Dataset Format (Simple + Practical) 📦

Each row corresponds to one clip directory with three assets:

• video (.mkv)
• srt_file_name (.srt)
• action_file_name (.jsonl)

In events.jsonl, each line includes:

• frame (frame index)
• t (seconds)
• analog controls (lx, ly, rx, ry, lt, rt)
• button state list (buttons)
• optional context fields like reset/mode metadata

This is designed so you can stream rows directly into action-conditioned training code with minimal preprocessing.

---

Why this is useful for world models 🧠

• Action-conditioned trajectories (not just passive video)
• Strong temporal alignment between controls and frames
• Mixed contexts (on-foot + vehicle segments)
• Collected for long-horizon, interactive modeling workflows

---

Quick Start (Python) 🚀

from datasets import load_dataset
import json
from huggingface_hub import hf_hub_download

ds = load_dataset("lucidml/GG800-Subset", split="train")
sample = ds[0]

# `video` is a Video feature object/handle.
video_obj = sample["video"]

# Sidecar files are stored as relative paths in the dataset repo.
srt_path = hf_hub_download(
    repo_id="lucidml/GG800-Subset",
    repo_type="dataset",
    filename=f"train/{sample['srt_file_name']}",
)
jsonl_path = hf_hub_download(
    repo_id="lucidml/GG800-Subset",
    repo_type="dataset",
    filename=f"train/{sample['action_file_name']}",
)

with open(jsonl_path, "r", encoding="utf-8") as f:
    first_event = json.loads(f.readline())

print("first frame event:", first_event)

---

License and Access 🔐

• GG800-Subset is released under the repository LICENSE file (research-focused terms).
• For larger subsets, enterprise access, or full-dataset discussions, contact:
  • abhishek@lucidml.ai
  • Subject: GG800 Access Inquiry

---

Citation 📝

If you use GG800-Subset in research, please cite the associated paper/technical report and mention:

"This work uses GG800-Subset, licensed under the GG800-Subset Research License v1.0."

---

Thanks for building with GG800-Subset.
If you create something cool, we would love to see it! 🌍⚡