Ethosoft/ck3-gameplay-mouse-keyboard-dataset

👑 CK3 GAMEPLAY MOUSE & KEYBOARD DATASET 👑

Every click. Every drag. Every panicked alt-tab before your heir gets assassinated.

"Domesticity is a trap." — also true of clean training data.

---

🏰 What is this madness?

Somewhere out there, a human is playing Crusader Kings III — scheming, marrying off third cousins, poisoning uncles, micromanaging a thousand vassals — and every single mouse move, click, scroll and keystroke is being recorded frame-perfectly alongside the screen itself.

This is not a "let's-play." This is a behavior-cloning goldmine: raw human-in-the-loop control data from one of the most UI-dense, decision-heavy grand strategy games ever made. If you've ever wanted to teach a model to actually play CK3 — navigate nested menus, drag sliders, double-click portraits, spam the council screen at 3am — this is the substrate.

Built with a custom recorder (CK3 Auto Dataset Recorder) that only captures input while CK3 is the active foreground window, so there's no keylogging of your Discord DMs hiding in here. 🛡️

🎬 What's inside

Every recording session is a self-contained folder, timestamped to the second:

ck3_20260602_210003/
├── screen.mp4           # 30 fps desktop capture (libx264, gdigrab)
├── input_events.csv     # every discrete event: move / click / scroll / key up-down
├── input_samples.csv    # uniformly sampled cursor + button + key state @ 60 Hz
├── metadata.json        # session metadata, sync info, exact ffmpeg invocation
└── ffmpeg.log           # full encoder log for the paranoid / the reproducible-research crowd

56 sessions are live on the Hub right now, streaming toward a full planned set of 187 recordings — roughly 66 GB of gameplay video plus ~1.8 GB of dense, millisecond-precision input telemetry across 374 CSV files. This dataset is actively growing — check back, it gets bigger while you're not looking, kind of like a CK3 empire under a competent ruler.

🖱️ The input data, unpacked

input_events.csv — every raw event, timestamped with nanosecond precision on two clocks (t_perf_ns, t_epoch_ns) plus a convenient t_sec:

event	fields
mouse_move	x, y
mouse_click	x, y, button, pressed
mouse_scroll	x, y, dx, dy
key_press / key_release	key

input_samples.csv — the same session, resampled to a clean 60 Hz grid: x, y, left, right, middle, pressed_keys. This is the one you want if you're training a policy and don't want to deal with irregular event timing.

metadata.json — the Rosetta Stone. FPS, sample rate, the exact ffmpeg command used, and — critically — the sync formula tying input timestamps to video frames:

video_time_sec ≈ input_t_sec − ffmpeg_popen_sec_from_t0

Sub-frame alignment between what the human saw and what the human did. No guesswork.

🧠 What can you build with this?

• Behavior cloning / imitation learning — train an agent that plays CK3 the way a human actually plays it: hovering, hesitating, misclicking, correcting.
• Video-to-action models — predict the next mouse/keyboard action from screen state, the GUI-agent holy grail applied to a genuinely brutal UI.
• Mouse trajectory & intent modeling — dense cursor telemetry at 60 Hz is rare and valuable well beyond games.
• RL bootstrapping — warm-start a reinforcement learning policy from real human demonstrations instead of random exploration into a 4X-strategy action space.
• HCI / UX research — study how humans actually navigate deeply nested menu systems under time pressure.

📂 Loading it

from huggingface_hub import snapshot_download

# grab everything (⚠️ ~66 GB and counting — you have been warned)
local_dir = snapshot_download(
    repo_id="Ethosoft/ck3-gameplay-mouse-keyboard-dataset",
    repo_type="dataset",
)

# or cherry-pick just one session's input data (kilobytes, not gigabytes)
from huggingface_hub import hf_hub_download
import pandas as pd

path = hf_hub_download(
    repo_id="Ethosoft/ck3-gameplay-mouse-keyboard-dataset",
    repo_type="dataset",
    filename="ck3_20260602_210003/input_samples.csv",
)
df = pd.read_csv(path)

💡 The .mp4 files are the expensive part. Point-in-time syncing means you can pull input_*.csv + metadata.json for cheap analysis, and only fetch screen.mp4 for the sessions you actually need frames from.

⚙️ Recording pipeline

• Capture: ffmpeg + gdigrab, full desktop, libx264 / veryfast / crf 20, 30 fps, cursor drawn in-frame
• Input hook: low-level mouse + keyboard listener, active only when ck3.exe has foreground focus (1s grace period on focus loss)
• Sync anchor: shared perf-counter epoch (t0_perf_ns) stamped at both video and input start

🎲 A note on the dataset's soul

This wasn't scraped, it wasn't synthesized, and no LLM pretended to click a button. It's a real person, real vassals, real broken alliances — grinding out actual gameplay while a recorder quietly wrote down every twitch of the wrist. Treat it with the reverence you'd give any dataset built from someone's genuine, slightly-too-many-hours hobby.

📜 Citation

@dataset{ck3_gameplay_mouse_keyboard,
  title  = {CK3 Gameplay Mouse Keyboard Dataset},
  author = {Yağız Ekrem Dalar},
  year   = {2026},
  doi    = {10.57967/hf/9535},
  url    = {https://huggingface.co/datasets/Ethosoft/ck3-gameplay-mouse-keyboard-dataset}
}

---

May your dynasty be long, your click precision higher, and your training loss ever-decreasing.