TOWER-SimData / VALIDATION.md
marcozzxx810's picture Hnyquister's picture
Prepare TOWER-SimData A-F paper-aligned release (#2)
7332d7b
|
Raw History Blame Contribute Delete
2.91 kB
# Release review scope
Reviewed data revision: `642a2bfbe1ca273f1d7e476187fde4e775db9e56` (PR #2).
This record distinguishes repository-level verification from sampled payload checks.
It does not certify every trajectory or replace the producer's qualification records.
## Independently checked
- Exactly 600 HDF5 paths, with 50 per A–F × 2/3-layer configuration.
- All manifest paths match the repository file inventory; no legacy Task3 HDF5 directories remain in this revision.
- All 600 published hashes match `CHECKSUMS.sha256` and Hugging Face's LFS object IDs; all sizes match LFS/repository metadata. This compares metadata, not 600 fully downloaded payloads.
- All `TASK_MAPPING.json` episode, frame and byte totals match the manifest: 1,659,079 frames and 129.92517953552306 GiB of HDF5 payloads.
- All 600 published hashes and all 600 source hashes are distinct.
- A-2 and C-2 samples were opened through HDF5 range reads. Start/middle/end JPEGs from all four cameras decoded at 320×240.
- One complete C-2 file was downloaded and SHA256-verified: `data/taskC-2_extract_and_stack_2/episode_000279_20261030T033940Z_e664b157.hdf5` (53,506,506 bytes). Its core state arrays were finite and its simulation timestamps strictly increasing. Twelve decoded images were visually inspected. The inspection loader was tested on this file.
- In the sampled C-2 file, checked root and object attributes contained no `/mnt/`, `/root/`, `/home/`, or `/Users/` paths. This is not an all-file privacy scan.
## Still needed from the release producer
1. Publish a sanitized source-pool manifest identity, exact deterministic selection rule, and held-out identity list or equivalent verifiable exclusion evidence. Bind each selected episode to that provenance. The current manifest records the selected files, but does not reconstruct selection from the claimed 728 training episodes or verify exclusion of the 24 held-out episodes.
2. Attach hash-bound per-episode qualification results for the released set, including the ordinary-grasp policy, approach/carry trajectory checks, and the scope of recorded-phase visual review. A success label, a reverse-grasp flag, or three decoded timestamps alone does not establish those checks.
3. Verify and document the training observation-to-action pairing, reference/issued/applied-command conversion and timing, and final-row policy. The raw inspection example deliberately does not choose a training convention.
4. Confirm image dimensions across all released files, finalize citation date when publication is approved, and preserve the distinction between this 600-episode SIM subset and the paper's full SIM/REAL dataset.
No complete video playback, all-episode checksum download, independent success rescoring,
or full-set posture requalification was performed in this review. Source recordings and
payloads have not been modified by the documentation follow-up.