p-doom/idm-eval-set

IDM Eval Set

A validation set for evaluating Inverse Dynamics Models on macOS screen recordings. Each sample is a 5-second clip of real desktop usage (browser, IDE, terminal) paired with a ground-truth action log captured at the OS level.

The task: given a short screen recording, predict the sequence of user input actions (keypresses, mouse clicks, scrolls) that produced the observed screen changes.

Dataset Structure

clips_recording_{uuid}_seg{N}/
  clip_000_{tag}.mp4       # 5s screen recording (1728x1080)
  clip_000_{tag}.json      # ground truth action log
annotations.json           # visibility labels per action
gt_overrides.json          # manual corrections to GT details

Stats

Clips	51
Recordings	11
Total raw actions	10,914
Resolution	1728 x 1080
Clip duration	5 seconds

Tag distribution:

Tag	Count
scroll/drag	18
keystroke-heavy	17
mixed	4
click-heavy	4
hotkeys	2
hard-case	1

Action Log Format

Each clip JSON contains:

{
  "start_s": 206.913,
  "end_s": 211.913,
  "tag": "keystroke-heavy",
  "actions": [
    [206933331, ["KeyPress", [32, "Space"]]],
    [207233331, ["KeyRelease", [32, "Space"]]],
    [208633331, ["MousePress", ["Left", 0, 0]]],
    [209533331, ["MouseScroll", [0, -1, 0, 0]]]
  ]
}

• Timestamps are absolute microseconds (subtract start_s * 1e6 for clip-relative)
• Action types: KeyPress, KeyRelease, MousePress, MouseRelease, MouseMove, MouseScroll, ContextChanged
• KeyPress params: [keycode, key_name]
• MousePress params: [button, x, y] (coordinates not captured in this version)
• MouseScroll params: [dx, dy, x, y]

Annotations

annotations.json contains manual visibility labels for each primary action (KeyPress, MousePress, MouseScroll) in each clip. Keys are clip paths, values map action indices to visibility labels:

Label	Count	Meaning
visible	485	Effect is directly visible in the frames
inferable	273	Effect can be inferred but isn't directly visible
ambiguous	25	Action type is unclear from video (e.g. scroll via mouse vs keyboard)
not_predictable	27	Cannot be predicted from video alone

{
  "clips_recording_.../clip_003_keystroke-heavy": {
    "0": "visible",
    "1": "inferable",
    "2": "ambiguous",
    "3": "not_predictable"
  }
}

Use these to filter ground truth when scoring — e.g. exclude not_predictable and optionally ambiguous actions from recall calculations.

GT Overrides

gt_overrides.json contains manual corrections to ground-truth action details (e.g. when a modifier key was held from before the clip). Structure:

{
  "clips_recording_.../clip_name": {
    "edits": {"5": "Cmd+Tab"},
    "deletions": [],
    "additions": [{"frame": 8, "type": "KeyPress", "detail": "Space"}]
  }
}

Apply overrides after processing raw GT through filter_gt_actions + coalesce_gt_events.

Gesture Evaluation (Mouse Movement + Scroll)

In addition to sparse event evaluation (KeyPress, MouseClick, MouseScroll), this dataset supports gesture evaluation: predicting per-frame mouse cursor movement and scroll magnitude.

Gesture GT Files

• gesture_gt_exp.json — Exponential bin indices (±1 to ±9 per axis)
• gesture_gt_norm.json — Normalized 0-1000 scale (resolution-independent)

Both are derived from the raw MouseMove and MouseScroll events in each clip JSON. Mouse deltas are accumulated per frame (5fps), normalized by video resolution, then binned (exp mode) or kept as integers (norm mode). Scroll sign convention: positive = scroll down (content moves up).

Format

• MouseMove details: "dx,dy" — signed bin indices (exp) or normalized integers (norm). Positive dx = right, positive dy = down.
• MouseScroll details: signed bin index or normalized integer. Positive = scroll down.

Exponential Bin Scale (mouse dx/dy)

Bin	Normalized range (0-1000)
±1	0-1
±2	1-3
±3	3-7
±4	7-16
±5	16-40
±6	40-95
±7	95-230
±8	230-550
±9	>550

Metrics

• R² (coefficient of determination): primary metric for mouse movement quality. VPT (OpenAI) reports R²=0.97 on their trained IDM.
• L2: Euclidean distance between predicted and GT bin vectors (lower = better).
• F1: frame-level detection (did the model predict the right frames?).
• DirAcc: direction accuracy (did dx/dy signs match?).

Evaluation