Oratis's picture
完成全量游戏采集归档与源清理核验,补齐逐文件哈希和固定分片索引
36551b3 verified
|
Raw History Blame Contribute Delete
4.47 kB
---
pretty_name: Diogenes Dying Light Gameplay Archive v01
license: other
license_name: original-gameplay-archive-terms
license_link: https://huggingface.co/datasets/DiogenesLab/Diogenes_Gameplay_Dying_Light_Archive_v01/blob/main/NOTICE.md
tags:
- video
- gameplay
- multimodal
- webdataset
- archival
size_categories:
- 10K<n<100K
---
# Diogenes Dying Light Gameplay Archive v01
Project-closeout public archive of **20,511 clips from 309 source
session directories**, preserving **123,374 original objects /
429,956,556,136 bytes**. Test recordings, failed QC, menus, cutscenes,
loading and capture defects are retained. Downloads require no account approval.
## Sample layout and reuse
Tar shards use WebDataset-style sample keys such as
`data/session0001-run0001-clip000001`. Public session/run/clip ordinals follow sorted
opaque source IDs, **not chronology**. Use original `session_id`, `segment_index`,
`session_frame_start`, frame timestamps, `fps_num/fps_den` and dimensions in metadata
for alignment. The same sample key groups the modalities below.
| Suffix | Contents |
| --- | --- |
| `.video.mp4` | Original H.264 game capture; may include menus, loading or faulty frames. |
| `.meta.json` | Capture configuration, target, timing, dimensions and audio health. |
| `.frames.jsonl.zst` | Zstandard JSON Lines of frame times/indexes, action/input states and capture health. |
| `.events.jsonl.zst` | Zstandard JSON Lines of input events; empty streams are preserved. |
| `.manifest.json` | Original logical-filename SHA-256 map. |
| `.audio.pcm.zst` | Present for 20,510 clips: headerless signed 16-bit little-endian PCM, 48 kHz stereo, Zstandard compressed. Original audio source is game-process loopback. |
Metadata schemas: `dm-raw-recording-1` (20,511 clips). Read fields from each file; do not assume schema
variants have identical fields or uninterrupted recording. `qc/sessionNNNN.acceptance.json`
preserves 309 reports: 0 passed and
309 failed. Missing audio remains missing; no silence file is invented.
`MANIFEST.jsonl` lists the exact byte size and SHA-256 of every public member.
Supporting metadata/QC files are not extra clips. Decompress `.zst` before parsing
JSONL or PCM. No train/test split or uniform action coverage is claimed.
```python
import json, tarfile
with tarfile.open("downloaded-shard.tar", "r:") as archive:
for member in archive:
if member.isfile() and member.name.endswith(".meta.json"):
meta = json.load(archive.extractfile(member))
print(member.name, meta["schema_version"], meta["segment_index"])
```
## Provenance, quality and rights
Historical metadata states:
- `non_trainable`: omitted in every clip; unknown, not inferred.
- `production_certified`: omitted in every clip; unknown, not inferred.
- `catalog.mode`: omitted in every clip; unknown, not inferred.
These facts, failed quality checks, empty inputs, dropped/duplicated frames and
other capture defects are preserved. Publication does not certify training
eligibility or grant rights to the underlying game IP; see `NOTICE.md`.
Source collection label: `dying-light-pc`.
All 82,353 structured metadata objects /
17,707,868 JSONL records were fully parsed and scanned.
Eighteen schema findings were reviewed from concrete source objects and capture-core input/projection code: numeric counts of discarded control-key events (virtual-key codes 119/121), Boolean capture-control markers and horizontal mouse-wheel fields. No private metadata findings remained.
Representative media review: 64 representative source sessions out of 309, using deterministic source-session ordering within the available capture/QC cohort. Gameplay, inventory/map/shop menus, NPC dialogue and faulty/black captures were observed; no personal account or desktop content was identified in those samples.
This is representative frame inspection, not exhaustive video viewing or semantic
audio transcription. Process loopback is the metadata-declared audio source.
Original objects were transferred without re-encoding. The complete reviewed collection is published and verified in 435 TAR shards.
`MANIFEST.jsonl` records each actual public member SHA-256 and byte size;
`ARCHIVE_INDEX.json` records every TAR hash, size and immutable verification revision.
Pinned Hugging Face LFS checks and anonymous range downloads passed before
corresponding R2 objects were removed. Capture/schema source:
https://github.com/CuddlerAI/Diogenes-Archive.