GY2233/open_p2p

Open-P2P 50-Sample Subset

A 50-sub-clip preview of the Open-P2P training data prepared for the wan22_action_train framework (Wan2.2 TI2V-5B + ActionModule, text-action conditioning).

This subset shares the exact same on-disk schema as the full pipeline output under /efs/yige/datasets/openp2p_shards_v2/. It is intended for quick inspection of the per-frame caption translation, the WebDataset-compatible schema layout, and the underlying video / action structure — without having to mount the multi-TB full corpus.

Provenance

• Source: Open-P2P dataset (108,738 user-recorded clips across 66 PC / Roblox games, 20 fps, 480p–720p captures, with binary annotation.proto per-frame action records).
• Conversion code: data_scripts/openp2p_to_shards.py from the upstream training repo, sliding-window-sliced to 81 raw frames @ 20 fps (= 4.05 s per sub-clip).
• Caption translation: per-frame keyboard / mouse / gamepad tokens are translated into short verb phrases using the per-domain semantic mapping under action_semantics/parts/group_{a..d}.yaml. Camera (mouse dx/dy) is intentionally left out of the caption channel — it is a numeric continuous signal, not a discrete user-issued action.

Files

Each sub-clip is stored as a flat triple under the dataset root:

<clip_name>.mp4              # 832x480 H.264, 20 fps, 81 frames (4.05 s)
<clip_name>.actions.json     # JSON array, length 81; per-frame schema below
<clip_name>.meta.json        # clip-level metadata
manifest.jsonl               # one row per sub-clip with extracted fields
README.md                    # this file

<clip_name> is <p2p_uuid>_<sub_clip_index_4digit>, e.g. 019a930b-1e13-7c90-9324-8443213e8175_0003.

actions.json schema (per frame)

{
  "frame":            0,
  "caption":          "move forward, dash, fire weapon",
  "keys":             ["w", "lshift"],
  "mouse_buttons":    ["left"],
  "gamepad_buttons":  [],
  "mouse_dx_dy":      [12, -5],
  "mouse_pos":        [640, 360]
}

meta.json schema

{
  "clip_name":        "019a930b-1e13-7c90-9324-8443213e8175_0003",
  "p2p_uuid":         "019a930b-1e13-7c90-9324-8443213e8175",
  "sub_clip_index":   3,
  "domain":           "DOOM Eternal",
  "num_frames":       81,
  "fps":              20,
  "src_fps":          20.0,
  "width":            832,
  "height":           480,
  "duration_s":       4.05,
  "start_frame":      243,
  "category":         "move forward"
}

Subset composition

• Total sub-clips: 50
• Distinct domains covered: 10
  • Counter-Strike 2, DOOM, DOOM Eternal, Left 4 Dead 2, Need for Speed, Quake, Roblox/blade-ball, Roblox/death-ball, Roblox/hypershot, Roblox/natural-disaster-survival
• Distinct dominant action categories: 14

Intended use

• Quick sanity check of the per-frame caption pipeline (raw key tokens → per-domain verb phrase).
• Reference fixture for the --use_latent_dataset / --action_cond_mode text training path in wan22_action_train.
• Offline visual inspection (mp4) paired with action overlays (actions.json).

License

Released under CC BY-NC 4.0 for non-commercial research only. Original Open-P2P recordings are user-contributed gameplay footage; respect the publishers' game ToS when reusing visual content.