Oratis's picture
Document reviewed capture archive, modalities, and historical quality limitations
ffbe644 verified
|
Raw History Blame Contribute Delete
4.11 kB
metadata
pretty_name: Diogenes Atomic Heart Gameplay Archive v01
license: other
license_name: original-gameplay-archive-terms
license_link: >-
  https://huggingface.co/datasets/DiogenesLab/Diogenes_Gameplay_Atomic_Heart_Archive_v01/blob/main/NOTICE.md
tags:
  - video
  - gameplay
  - multimodal
  - webdataset
  - archival
size_categories:
  - n<1K

Diogenes Atomic Heart Gameplay Archive v01

Public project-closeout archive: 1 recorded sessions, 39 clips, 387.733 seconds and 11,632 frames. The complete 235 source objects total 777,241,049 bytes. No account or access request is required to download. Historical failed QC and non-trainable status are retained; this is an inspection/reproducibility archive, not certified training data. No game binaries or downloaded model weights are included.

Sample layout and schema

Each clip uses one WebDataset-style sample key, for example data/session0001-run0001-clip000000, with six field suffixes:

Suffix Meaning
.video.mp4 Original H.264 gameplay video.
.audio.pcm.zst Zstandard-compressed headerless signed 16-bit little-endian PCM; see each meta file for sample rate/channel count. Recorded source is game-process loopback, not microphone.
.frames.jsonl.zst Per-frame timestamps/indexes, key/button/action states, pointer movement and duplication/drop markers.
.events.jsonl.zst Timestamped input events; may be empty.
.meta.json dm-local-dev-recording-2 schema: game/catalog, encoder/audio parameters, stream hashes and capture health.
.manifest.json Original SHA-256 map keyed by logical field names such as video.mp4.

qc/sessionNNNN.acceptance.json retains each session's full historical QC report. MANIFEST.jsonl is the public member-size/SHA-256 index. These are metadata, not extra clips; filter tar members under data/ when iterating samples. Decode .zst bytes with Zstandard before reading JSON Lines/PCM. frame indexes are clip-local; session_frame preserves the continuous frame sequence. Frame rate and dimensions are recorded in each clip's metadata. No curated train/test split is claimed.

A dependency-free metadata reader (no path extraction):

import json, tarfile
with tarfile.open("downloaded-shard.tar", "r:") as archive:
    samples = {}
    for member in archive:
        if member.isfile() and member.name.startswith("data/"):
            key, field = member.name.split(".", 1)
            samples.setdefault(key, {})[field] = member.name
    sample = samples[sorted(samples)[0]]
    metadata = json.load(archive.extractfile(sample["meta.json"]))
    print(metadata["n_frames"], metadata["fps_num"], metadata["fps_den"])

Provenance, quality and verification

Every clip retains non_trainable: true, production_certified: false, and catalog.mode: dev_non_trainable. All 1 original acceptance reports failed QC. Consult their check-level values before interpreting inputs or training any model; missing/empty inputs and duplicate frames have not been corrected. Publication does not establish training authorization or game-content rights.

All 157 structured metadata files (24,395 JSONL rows) were read and scanned for account/email/token identifiers and private paths. Generic capture-device profiles and opaque session/target identifiers remain. Representative video frames and MP4 container tags were inspected across every session. This is sampling, not exhaustive visual annotation or semantic audio review. Audio source provenance is the recorded process-loopback setting.

Original object bytes are preserved; archive member names are normalized to public sample keys. The public index omits private storage and operator-machine paths. Metadata digests were checked during review; all media/source digests must match the original per-clip manifest during transfer. A shard is verified against pinned Hugging Face LFS size/SHA-256 before any source is eligible for deletion.

Capture/schema code: https://github.com/CuddlerAI/Diogenes-Archive. See NOTICE.md for original content rights and historical eligibility constraints.