Gameplay Alliance

About

5,000 hours in the archive. £1,500 buys a 100-hour entry block at £15 per hour. Larger volumes and production to specification are quoted separately — see Pricing below.

About

Gameplay Alliance produces PC gameplay data to specification: you define the volume, the titles and the capture spec, and we produce it. 5,000 hours are available from archive today, and capture runs continuously at approximately 2,000 additional hours per week.

Every session pairs video with frame-level input on a common session clock. Most gameplay datasets are video only: you see what happened, not what the player did. Here the input that produced each frame ships alongside it, so a model can learn the relationship between a decision and its consequence rather than the appearance of a scene.

Nothing here is repurposed entertainment footage. There is no commentary overlay, no editing, no post-production — encoder and capture settings are consistent because the capture software is ours. In-game graphics settings are set by the contributor and are not collected on archive sessions.

Pricing

Priced per hour of gameplay. Minimum order: 100 hours. The listed price is that entry order at our standard rate of £15 per hour.

The rate applies to the existing archive: sessions already captured, licensed as they stand. Specific titles or genres may carry a different rate, and rates for large committed volumes are agreed case by case.

Production to specification is quoted separately and prices above the listed rate. You define the titles, the volume and the capture spec, and we commit capacity against it. Semantic annotations, instrumented synchronisation, device calibration, custom telemetry formats and human transcripts are scoped within that quote.

Tell us what you need and we will quote it.

Dataset Features

Per session

Video with the game audio mix. No microphone, no voice chat Raw input events: keyboard, mouse buttons, wheel, and raw mouse deltas, timestamped on the session clock Per-frame input state on a fixed timeline at the encoded frame rate, with input aggregated over each frame interval Per-frame condition label — new present, repeated present or capture missing — inferred visually, with its confidence declared per session Keybind mapping per title, split by control context, with per-binding provenance and declared coverage Title and session identification, build identifier Session timing and duration, resolution and frame rate Display and encoder configuration; full host hardware on production deliveries Creator country Clock-sync mapping with its measured error, declared per session SHA-256 checksums for every media and telemetry file Per-session provenance record covering consent, capture date and protocol

Available as custom work

Semantic annotations to your schema · instrumented synchronisation · device calibration · control-context stream · episode boundaries and gameplay events.

Distribution

5,000 hours available from archive today, single-player and multiplayer Approximately 2,000 additional hours per week, produced to specification Video: MP4 (H.264, progressive, faststart) Metadata: structured JSON. Input telemetry also available as CSV or Parquet Per-title volumes available under NDA. Delivered size for your scope confirmed on quotation Collection Method

Gameplay Alliance is a community of registered creators — approximately 10,000 active gamers across 18 countries in Latin America — recording on capture software we built and operate. We do not source through third-party studios or aggregators.

Consent is secured at the source: contributors accept explicit data-licensing terms before recording, not retrospectively, and every session carries its own provenance record naming the consent basis and its version. Consent is revocable; on withdrawal we notify recipients of record and the affected sessions are removed from the delivered set.

For the underlying commercial titles we conduct a title-by-title review and exclude any title whose terms explicitly restrict the use of gameplay for AI training. Where a title requires it, we hold agreements with regional publishers covering the material captured.

Every session passes two review passes, automated and then human, covering video integrity, input synchronisation, inactive content, personal information on screen and conformance to the agreed specification. The conformance report ships with the data and declares the result of every check, including any that did not pass. Sessions that fail the specification agreed for your order do not reach delivery.

Usage

World-model training and evaluation · action-conditioned video generation · behaviour cloning and imitation learning · agent benchmarking against human baselines · intent and task-segmentation research.

Coverage

Licence coverage: worldwide. Use is not restricted by territory Capture and supply: Latin America, across 18 countries Platform: PC Languages: primarily Spanish, with Brazilian Portuguese and English available Transcripts: human-created when required. We do not use AI-generated transcripts Delivery

Delivered by direct transfer to a cloud storage bucket you nominate, S3 by default, using the same pipeline we run for our enterprise clients. After purchase we confirm scope, agree the destination and the folder structure, and transfer in batches. Every batch ships with a manifest listing each session, its duration, its format and its checksum, so you can verify what arrived against what was ordered.

License

Proprietary, licensed by Gameplay Alliance (Sunrise Advisors Generation Ltd) under the Opendatabay AI Training and LLM Use License permitting commercial use: you may train, fine-tune and evaluate models on this data and deploy the resulting models commercially, including in production and API services. The dataset itself may not be redistributed or resold.

Perpetual or term, structured to your requirement. Exclusive arrangements available on request.

Known limitations

These apply to the existing archive. Where there is a route to resolution, it is noted.

Synchronisation is estimated, not instrumented, on existing sessions. Each session declares its own figures. In a representative session the cross-correlation bound is 33 ms, on top of a declared mapping uncertainty of 50 ms and a start-anchor spread of up to 110 ms measured across machines. Because the cross-correlation includes the title's own input-to-photon latency, it is an upper bound rather than a measured offset. Instrumented synchronisation is available as custom work. Frame condition is inferred visually on the full-resolution decode rather than read from a source present identifier, which OS-level capture does not expose. Confidence is declared per session. Repeated presents occur when the capture rate is not matched by the title's render rate or by the host's display mode. The fraction is measured and declared per session; in a representative session it is 9.5 per cent.

Mouse calibration was not collected on existing sessions: DPI, pointer settings and in-game sensitivity are absent, so raw counts cannot be converted to degrees of view rotation without calibration. Available as custom work. Bindings are ours, not the contributor's. The key map is our curated map per title, split by control context; personal rebinds are not reflected, and keys pressed outside the curated map are flagged rather than labelled. The map is compiled from published control references and is not verified in-game; per-binding provenance ships in the device descriptor. No control-context stream is delivered, so which context applies at a given moment must be inferred from the video. All of this is addressable as custom work. Title state, camera and entity transforms, coordinate frames and episode boundaries cannot be derived from screen capture of a title we do not control, and are not part of this product. Capture is at OS level, not inside the title. The title's own render resolution, graphics preset, FOV and render scale are not known, and pointer-capture state is not populated on archive sessions. Keyboard codes come from the Windows virtual-key code. On non-US layouts, punctuation keys may not correspond to the physical position. A small fraction of timeline slots can be missing where capture dropped. The fraction is measured and declared per session; in a representative session it is three in one hundred thousand. Distribution across titles and genres is uneven — breakdown available under NDA. Sample

A complete session is available as a free technical sample: video, raw input events, per-frame input state, session manifest, capture configuration, device descriptors and our QA self-assessment. Provided for evaluation of format and synchronisation; not licensed for model training.

Some fields carried on production deliveries are not populated in the free sample, among them full host hardware and creator country. The sample's own manifest lists what is partial, and its conformance report declares every check including those that did not pass.