p-doom/idm-eval

IDM Eval Set

A validation set for evaluating Inverse Dynamics Models on macOS screen recordings. Each sample is a 5-second clip of real productivity desktop usage (browser, IDE, terminal, docs, dashboards) 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, cursor moves) that produced the observed screen changes.

Code

Training, inference, and evaluation code is available at p-doom/inverse-dynamics-model.

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
gesture_gt_norm.json       # per-frame mouse-move / scroll GT (0 to 1000 scale)
gesture_gt_exp.json        # per-frame mouse-move / scroll GT (exponential bins)

Stats

Metric	Value
Clips	44
Recordings	10
Total raw actions	8,504
Resolution	1728 x 1080
Clip duration	5 seconds

Tag Distribution

Tag	Count
scroll/drag	15
keystroke-heavy	15
click-heavy	6
hotkeys	4
mixed	3
hard-case	1

Baseline leaderboard

The table below provides reference scores for p-doom/idm and off-the-shelf VLM baselines on this eval set. Scores use visibility filtering (visible + inferable) and action-type-specific matching. MM R² and MM cos_mean include missed MouseMove frames as zero predictions.

Model	Overall F1	KeyPress F1	MouseClick F1	MouseMove F1	MouseScroll F1	MM R²	MM cos_mean	MM cov.
Ours (8B)	0.787	0.791	0.598	0.857	0.447	0.708	0.643	92%
Gemini 3.5 Flash	0.740	0.826	0.726	0.760	0.337	0.714	0.560	64%
GPT 5.5	0.709	0.821	0.714	0.669	0.392	0.586	0.455	52%
Kimi K2.6	0.540	0.711	0.444	0.381	0.326	0.420	0.177	25%
Gemma 4 31B	0.430	0.381	0.581	0.500	0.237	0.077	0.228	37%
Qwen3-VL 8B	0.360	0.409	0.449	0.334	0.127	-6.038	0.035	28%

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]]]
  ]
}

Details:

• Timestamps are absolute microseconds. Subtract start_s * 1e6 for clip-relative.
• Raw event types in the JSON: KeyPress, KeyRelease, MousePress, MouseRelease, MouseMove, MouseScroll, ContextChanged.
• For IDM evaluation, the canonical action set is KeyPress, MouseClick, MouseScroll, MouseMove. A MouseClick is derived from a MousePress event (its MouseRelease partner is discarded).
• KeyPress params: [keycode, key_name].
• MousePress params: [button, x, y]. The x, y fields are always 0 in this version (cursor position is not stored on press events); only the button name is meaningful.
• MouseScroll params: [dx, dy, x, y].

Annotations

annotations.json contains manual visibility labels for each primary action (KeyPress, MousePress, MouseScroll) in each clip.

Label	Count	Meaning
visible	479	Effect is directly visible in the frames
inferable	21	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	24	Cannot be predicted from video alone

Format:

{
  "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).

The eval set is strongly visible-dominated: the vast majority of annotated actions have a directly observable visual effect, so a competent IDM should be able to recover them from pixels alone without context inference.

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"}]
  }
}

Gesture GT (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_norm.json: normalized 0 to 1000 scale, resolution-independent.
• gesture_gt_exp.json: signed exponential bin indices (one bin per magnitude order).

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 either kept as integers (norm) or binned (exp).

Format:

• MouseMove details: "dx,dy". Positive dx = right, positive dy = down.
• MouseScroll details: signed scalar. Positive = scroll down (content moves up).

Scope

This eval set is productivity-focused. It covers IDE, browser, terminal, docs, dashboards, and PDF reading workflows. Gaming and entertainment clips are intentionally excluded so the benchmark targets long-horizon digital-work behavior.