Oratis's picture
Finalize verified depth probes with complete source hashes and fixed archive revisions
8e5e76d verified
|
Raw History Blame Contribute Delete
3.66 kB
---
pretty_name: Diogenes Fallout4 Historical Depth Probe Archive v01
license: other
license_name: original-capture-archive-terms
license_link: https://huggingface.co/datasets/DiogenesLab/Diogenes_Fallout4_Depth_Probe_Archive_v01/blob/main/NOTICE.md
tags:
- depth
- archival
- non-trainable
size_categories:
- 10K<n<100K
---
# Diogenes Fallout4 Historical Depth Probe Archive v01
The project has closed. This owner-authorized archive preserves **15,447 raw depth frames plus one global JSONL index / 56,946,101,409 source bytes** from `depth-retry-20260805-134854`. **All listed objects are fully archived and verified.** See [ARCHIVE_INDEX.json](ARCHIVE_INDEX.json) for every pinned tar revision, size and SHA-256. These are historical diagnostic captures, not accepted delivery data or newly certified training material.
## Format and limitations
Each `.f32` file is exactly 1280 × 720 × 4 bytes, little-endian IEEE-754 float32 in row-major order. Values represent the captured device depth buffer, **not linear metres**. The original index identifies source format `d24s8` and contains `present_seq`, dimensions, the generic DSV profile, relative file path and readback latency. Every one of the 15,447 index rows maps to exactly one stored frame; no indexed frame is missing from the reviewed source inventory.
The historical Fallout 4 inspection reported traditional depth (clear value 1), about 11.5% sampled empty/menu/loading frames, and 1–2 frames of readback latency. Calibration was not independently accepted. A later engineering decision marked this batch unusable for its intended delivery and discussed FP16 precision loss beyond about 20 m. The archived objects themselves are 32-bit float containers and their index labels the original source `d24s8`; the archive preserves that distinction and does not invent a corrected precision or metric interpretation. The first and last inspected frames are all-clear values. No paired RGB, audio or input stream is present in this prefix.
The archive keeps every original frame, including blank or failed examples; no metric reconstruction, resampling, pixel repair, status promotion or flag backfill was performed. Original project analysis documents are historical evidence, not a new calibration certificate. No absent `non_trainable` or `production_certified` field has been fabricated. See `NOTICE.md` for third-party rights.
## Reading and integrity
Uncompressed TAR shards under `archive/` keep consecutive groups of at most 250 complete frame files (about 922 MB). The global `depth-index.jsonl` is retained in a separate index shard. Every tar ends in its own `MANIFEST.jsonl` containing exact SHA-256 and size. The repository manifest now records every original source SHA-256, size and containing tar at a fixed revision. All pinned Hugging Face LFS hashes/sizes and anonymous downloads were verified before the original storage keys were removed.
Read frames with `numpy.frombuffer(bytes, dtype='<f4').reshape(720,1280)`; interpret the stored depth convention and calibration limits before using numerical values. Original paths are unchanged, so index-relative links resolve after collecting the shards under the same prefix.
The entire JSONL index was scanned in the cloud and read locally with matching SHA-256. All object sizes and index paths were checked; first, middle and last frame bytes were inspected as numeric buffers. All raw frame SHA-256 hashes were computed and verified during transfer. Review does not imply semantic quality acceptance or new game-content permissions.
Capture code: [Diogenes-Archive](https://github.com/CuddlerAI/Diogenes-Archive).