Oratis's picture
归档历史采集元数据与训练笔记,保留原始资格矛盾和固定视频出处
21fe504 verified
|
Raw History Blame Contribute Delete
3.94 kB

Historical capture metadata and training notes

This supplement preserves 1,716 distinct payloads: 847 raw-input action JSONL files, 857 capture metadata JSON files, one corrected Kimi Linear sweep result, ten research scripts/logs, and one historical WM01 video-dataset declaration. Logical aliases and original/public SHA-256 values are listed in MANIFEST.jsonl. The 857 associated full capture videos were independently hashed and are already preserved at the fixed public revisions in VIDEO_LINEAGE.jsonl. The raw action JSONL files and capture metadata retain their exact source bytes; generic run/session identifiers, timing anchors, physical key vocabulary, actions, numeric results and provenance are unchanged.

Historical status and limitations

The capture metadata is internally inconsistent: an owner-attested validation layer sets first_party=true, rights=first_party, non_trainable=false and is_trainable=true, while its tags retain no_publish and no_product_training; its original capture provenance still says non_trainable_complete and dev_non_trainable. These are historical source assertions, not an independent certification of rights, training eligibility or production quality. The project owner's retirement request authorizes this public archival copy; it does not erase the recorded restrictions or confer rights to third-party game content. production_certified=false is retained. Do not treat an attestation string as a separate legal release.

The old dataset declaration records 1,443 clips, 93 frames, 16 fps and 832×480 pixels. It is preserved unchanged as a historical declaration, separately from the actual MP4 metadata and visual review of those distinct older video variants. The older variants must not be relabeled as the previously archived 24-fps dataset. Existing caption/action files that match the public 24-fps archive by exact bytes can be represented by explicit member aliases; that match does not prove the different video encodings or timing are identical.

The research files preserve the corrected Kimi Linear decode-step sweep, original benchmark/sweep logs, WM01 baseline and post-arm launchers, original and revised action-conditioned rollout code, the pairing repair, and the Phase2 LongLive patch. They were statically inspected and parsed where applicable; none was executed for this archive. Original logs retain kernel compilation/autotuning, shutdown warnings, zero-valued memory readings and measured values. The corrected sweep records 64 decode steps and contexts up to 262,144; these are historical observations, not a newly reproduced performance claim. The rollout code explicitly documents blockwise temporal-attention approximations and its batch/frame constraints. Pairing repair and checkpoint/arm provenance remain intact.

Only a personal home-directory prefix and a single internal network endpoint appearing in two logs were minimized. The endpoint uses the documentation address 192.0.2.1; loopback addresses, generic system paths, software versions, model names, public documentation links and numerical scientific results are unchanged. Private originals and the exact replacement ledger remain retained. An archival credential-lookup script and two installer/download logs remain private operations records. They contain no additional scientific payload absent from these retained scripts/results.

The normalized TAR has neutral owner/timestamp headers. Research records retain this archive's CC BY-NC 4.0 terms; Diogenes-authored code follows the archived project's Apache-2.0 terms. LongLive-related patch/context code retains upstream attribution and Apache-2.0 obligations; see the unchanged LICENSE already archived under retired-training-source/kda-history/historical-source.tar.gz at LongLive/LICENSE and https://github.com/NVlabs/LongLive. Third-party game/model rights and historical input-use restrictions are not expanded.