Access is restricted. Enter the access code supplied to you by Diogenes Lab. Requests submitted without a valid code are rejected. This dataset must not be used to train, fine-tune, or evaluate any model.

Diogenes Gameplay Multigame Sample v01

Human gameplay recordings across 19 PC titles, captured with the same recording pipeline and the same per-frame action-annotation schema on every title. The point of this sample is breadth: one session per title, identical schema, per-title semantic action vocabularies (mappings/), and honest per-segment QC columns computed from the shipped data itself.

19 sessions · 603 segments · 1.72 h · 205,091 annotated frames · 332,572 raw input events
Video: 1920×1080 H.264, 11.47 GiB. Audio: 48 kHz s16le stereo (zstd), 0.74 GiB.
Segments are gapless 10 s storage units of one continuous recording per title (session_frame increases across segment boundaries, so segments concatenate losslessly). One title uses 60 s units; two titles are captured at 60 fps.
Every record: non_trainable: true, production_certified: false. See Governance.

Titles

Title	`game_id`	genre	view	segs	min	fps	sustained-input (med)	input balance (med)	audio	flagged segs
Atomic Heart	`atomic-heart-pc`	action role playing	first person	39	6.5	30	0.70	0.12	—	39
Avatar: Frontiers of Pandora	`avatar-frontiers-of-pandora-pc`	action adventure	first person	32	5.3	30	1.00	0.58	✓	2
Call of Duty® Steam（release 测试，禁止训练）	`call-of-duty-steam-release-test`	first person shooter	first person	32	5.3	60	1.00	0.44	✓	7
Cyberpunk 2077 Steam（本机 RGB 预检，禁止训练）	`cyberpunk-2077-steam`	action role playing	first person	33	5.3	30	1.00	0.61	✓	9
DEATHLOOP	`deathloop-pc`	first person shooter	first person	33	5.4	30	1.00	0.60	—	33
DOOM Eternal	`doom-eternal-pc`	first person shooter	first person	33	5.5	30	1.00	0.56	—	33
Dying Light	`dying-light-pc`	survival horror	first person	35	5.8	30	1.00	0.48	✓	2
Euro Truck Simulator 2 (Steam, in-cab driving, dev non-trainable)	`euro-truck-simulator-2-steam`	driving simulation	first person	33	5.3	30	0.90	0.00	✓	33
Fallout 4 Steam（release 测试，禁止训练）	`fallout4-steam-release-test`	action role playing	first or third person	6	5.2	30	1.00	0.57	✓	1
Far Cry 5	`far-cry-5-pc`	first person shooter	first person	32	5.3	30	1.00	0.55	✓	1
Forza Horizon 5	`forza-horizon-5-pc`	racing	first or third person	34	5.5	30	1.00	0.00	✓	34
Grand Theft Auto V Legacy	`gta5-legacy-pc`	open world action adventure	first or third person	33	5.4	30	1.00	0.36	✓	10
Halo Infinite Steam（release 测试，禁止训练）	`halo-infinite-steam-release-test`	first person shooter	first person	32	5.3	60	1.00	0.66	✓	2
Indiana Jones and the Great Circle	`indiana-jones-great-circle-pc`	action adventure	first person	33	5.4	30	1.00	0.41	✓	5
Prey (2017)	`prey-2017-pc`	action role playing	first person	35	5.8	30	1.00	0.56	✓	1
Red Dead Redemption 2	`red-dead-redemption-2-pc`	action adventure	first or third person	32	5.2	30	1.00	0.44	✓	9
Resident Evil Village	`resident-evil-village-pc`	survival horror	first person	35	5.8	30	1.00	0.58	—	35
Starfield	`starfield-pc`	action role playing	first or third person	29	4.7	30	1.00	0.46	✓	5
Titanfall 2	`titanfall-2-pc`	first person shooter	first person	32	5.3	30	1.00	0.68	✓	5

sustained-input = fraction of 1 s windows with any input activity. input balance = min(keyboard-active frame fraction, pointer-active frame fraction); driving titles are keyboard-steered, so their pointer fraction — and therefore balance — is legitimately 0. flagged segs counts segments carrying any qc_flags entry (see QC below).

Known limitations (read before evaluating)

Silent audio on 4 titles (atomic-heart-pc, deathloop-pc, doom-eternal-pc, resident-evil-village-pc): the process-loopback capture recorded digital silence for these titles in this batch. Video and input streams are unaffected. Audio files are shipped anyway so the failure mode is inspectable; every affected segment carries the silent_audio flag.
Keyboard-only titles (euro-truck-simulator-2-steam, forza-horizon-5-pc): driving titles were played without mouse; their segments carry one_sided_input.
10 segments with zero input events (menus/idle), each flagged no_input — kept because a breadth sample should show what non-interactive stretches look like in this schema, not hide them.
Final segments of a recording can be short and duplicate-heavy (short_tail_segment, high_duplicate): frames freeze while the recording is being stopped. This is expected stop behavior, not sensor failure.

QC

Per-segment QC columns in segments.parquet (all recomputed at build time from the shipped frames/events rows and decoded audio, never hand-entered): qc_sustained_input, qc_input_balance, qc_keyboard_frame_frac, qc_pointer_frame_frac, qc_action_frame_frac, qc_duplicate_frac, qc_audio_peak, qc_audio_rms, qc_flags.

Flag semantics: no_input (zero device events), low_sustained_input (sustained < 0.8), one_sided_input (balance < 0.2), high_duplicate (>10 % duplicate frames), short_tail_segment (<90 % of nominal segment length), silent_audio (peak ≤ 2 LSB), audio_degraded (capture health not ready; one segment, a 14 ms silence-filled discontinuity).

Integrity: every shipped file was verified against its per-segment SHA-256 manifest before upload; frame counts agree three ways (container probe, per-frame annotation rows, run manifest) on all 603 segments.

Layout

data/segments.parquet   one row per segment (index + QC + provenance)
data/frames.parquet     one row per video frame: keys, actions, pointer deltas
data/events.parquet     raw device events with timestamps
videos/<game_id>/<session>/<run>/<segment>.mp4
audio/<game_id>/<session>/<run>/<segment>.pcm.zst   48 kHz s16le stereo
mappings/<game_id>.json per-title semantic action vocabulary + genre taxonomy
qc/qc_segments.json     the QC table as plain JSON

Time model

Every frame and event carries three clocks: qpc (authoritative 10 MHz monotonic tick, run-scoped; frames carry their aggregation window output_qpc_start/end, events their own qpc — attribution is qpc ∈ [output_qpc_start, output_qpc_end)), video_t_us (segment-relative µs, the mp4's own coordinate), run_t_us (run-relative µs), and t_utc (anchored to a wall-clock timestamp recorded at run start; unlike earlier Diogenes samples this is a recorded anchor, not an estimate).

Actions are decoded per title

frames.actions strings come from each title's own binding (mappings/<game_id>.json): the same schema yields combat.primary_attack on one title and vehicle.throttle on another. Comparing two titles' segments side by side is the intended use of this sample.

Governance

Same terms as the other Diogenes Lab gated samples: access is by request with an access code; every row carries non_trainable: true and production_certified: false. This data must not be used to train, fine-tune, or evaluate any model. It is provided solely for data-quality and schema evaluation. Redistribution is not permitted.

Event type counts: mouse_move 290,689 · key_down 33,241 · key_up 6,030 · mouse_down 1,292 · mouse_up 1,280 · wheel 40.

Downloads last month: 7

Total file size:

13.1 GB