LongWorld training inputs
Historical training records and recovery mappings. Incomplete checkpoint weights, training-state files and superseded source-data archives were retired under the final-model-only policy.
Final inference exports are in LongWorld-Checkpoints; the selected historical training data now reside in the private prepared-data dataset. Evaluation results and handoff documentation are indexed in the private research archive.
Contents
| Path | Files | Size | Content |
|---|---|---|---|
training/ |
44 | 29,325,985 bytes | Small historical run logs and configurations; incomplete checkpoint directories have been retired. |
Files in this archive require manual access approval. The repository card and metadata remain visible. Candidate, local-probe, partial-run, and other original evidence labels remain unchanged.
Private research archive
Evaluation results (evaluations/), training-process reports (worktree/), and handoff and
archival documentation (documentation/) are retained in the private research archive
Xnhyacinth/LongWorld-Research-Private under training-state/, preserving their original
relative paths. The root-level recovery map (CLEANUP_20260930_RESTORE_MAP.json) is retained
there as well under training-state/. Access on request.
Related public resources
- LongWorld-Data: the 542-row local-engineering workflow release.
- LongWorld-128K-Baselines: ACC, LongTrace, and LongMIT inputs.
- LongWorld-Checkpoints: final step-680 checkpoint variants with loading files.
Historical documentation
The original README and original repository snapshot retain the detailed historical layout, manifests, and recovery records under the current access controls. Earlier references to public access describe the historical state. Historical Git records remain, but some duplicate historical LFS payloads were retired in the migration described below; affected historical download links no longer work.
Historical duplicate migration β 2026-10-30
After user approval, this repository retired 37 historical LFS objects from this repository after verifying byte-identical copies in the private research repository. Current main-tree files are outside this migration. Historical pinned revisions and tags that reference the retired LFS objects may no longer download those bytes, and existing scripts pinned to those paths need to use the mapping. The retained copies require private-repository access. See the migration map and resolver for scope and completion status.
Additionally, 48 historical RNG objects and four scheduler objects belonging to incomplete old resume states were retired. Cleanup record.
Historical benchmark duplicate cleanup β 2026-10-30
Eight historical GraphWalks/MRCR parquet objects match pinned public upstream datasets; all upstream bytes have been downloaded and SHA-256 checked. Restore map, script and status. The existing source-revision snapshot downloader is already incomplete from prior migrations; use the restore mapping for these historical paths.
Final-model-only cleanup β 2026-10-30
The incomplete LongTrace Instruct step-200 model shards, intermediate checkpoint directories, historical training-argument binaries, and redundant tokenizer copy were retired at the ownerβs request. The final step-680 model exports and P456 inference export are outside this cleanup. Historical snapshot manifests may still describe retired paths. Cleanup receipt. The two historical SFT archives were retired after their selected training data and provenance were verified in the private prepared-data dataset.
Historical prepared data migration β 2026-10-30
The historical Swift training inputs and selected legacy data are now independently available in private LongWorld-Prepared-SFT-Data. Migration mapping and completion status. Public sanitized ACC/LongTrace/LongMIT remains in LongWorld-128K-Baselines.
- Downloads last month
- 1,140