README.md

---
license: cc0-1.0
task_categories:
- reinforcement-learning
- robotics
- image-to-video
- image-text-to-video
- image-to-3d
language:
- en
tags:
- world-model
- reinforcement-learning
- human-in-the-loop
- agent
pretty_name: No Man's Sky High-Fidelity Human-in-the-loop World Model Dataset
size_categories:
- 100K<n<1M
---

# No Man's Sky High-Fidelity Human-in-the-loop World Model Dataset

## Overview
This dataset is designed for **world model training** using real human gameplay data from *No Man’s Sky*.  
It captures **high-fidelity human–computer interaction** by recording both the game video and time-aligned input actions, preserving the realistic latency characteristics of a human-in-the-loop system.

Compared with “internal game state” datasets, this dataset retains the **physical interaction chain** (input → game/render → screen → capture), making it well-suited for training models that need to operate under real-world latency and sensory constraints.

## Dataset Structure
Each recording session is stored in a UUID directory.  
A typical session contains:
<UUID>/
  recording.mp4
  actions.jsonl
  events.jsonl
  metadata.json
  actions_resampled.jsonl

### 1) `recording.mp4`
The recorded gameplay video.  

### 2) `actions.jsonl` (per-frame input state)
One JSON object per video frame. Each entry contains the input state sampled at frame time.

**Schema:**
- `frame` (int): frame index  
- `timestamp_ms` (int): wall-clock timestamp in milliseconds  
- `frame_pts_ms` (float): frame time in milliseconds (PTS-based)  
- `capture_ns` (int): OBS compositor timestamp in nanoseconds  
- `key` (string[]): list of pressed keys at this frame  
- `mouse` (object):
  - `dx` (int): accumulated mouse delta X during the frame
  - `dy` (int): accumulated mouse delta Y during the frame
  - `x` (int): absolute mouse X position
  - `y` (int): absolute mouse Y position
  - `scroll_dy` (int): scroll delta during the frame
  - `button` (string[]): pressed mouse buttons (e.g., `LeftButton`, `Button4`)

### 3) `events.jsonl` (raw sub-frame input events)
Raw input events with microsecond timing, captured from the OS event stream.

**Schema:**
- `type` (string): event type  
  - `key_down`, `key_up`, `flags_changed`
  - `mouse_move`, `mouse_button_down`, `mouse_button_up`
  - `scroll`
- `timestamp_ms` (int): wall-clock timestamp  
- `session_offset_us` (int): microsecond offset from session start  
- `key` (string): key name for key events  
- `button` (string): mouse button name  
- `dx`, `dy`, `x`, `y` (int): mouse movement  
- `scroll_dy` (int): scroll delta  

### 4) `metadata.json`
Session-level metadata and video info.

**Schema:**
- `stream_name` (string): session UUID  
- `game_name` (string): game name  
- `platform` (string): `mac` / `windows` / `linux`  
- `video_meta` (object):
  - `width` (int)
  - `height` (int)
  - `fps` (float)
  - `total_frames` (int)
  - `duration_ms` (int)
- `input_latency_bias_ms` (number): recommended latency bias for alignment

### 5) `actions_resampled.jsonl`
High-precision resampled per-frame actions reconstructed from `events.jsonl` using latency compensation.  
This is the recommended aligned input stream for training.

---

## Suggested Usage
- For **world model training**, use `recording.<ext>` + `actions_resampled.jsonl`.
- For **analysis or recalibration**, use `events.jsonl` and `metadata.json`.

---

## Notes
- The dataset captures realistic system latency; alignment is provided but does **not** remove physical pipeline delay.
- This design targets **high-fidelity human-in-the-loop interaction** for robust world-model learning.