Hibiki research artifacts — Seeker38/hibiki_temp

Public research-results transfer requested by the project owner on 12 November 2026. This preserves historical Hibiki implementation/evaluation artifacts for continuation on another server. It is not a benchmark dataset, a new experiment, a claim of method superiority, or a replacement for the original server's restricted data archive.

Upload status: read DATASET_STATUS.json. Until that file exists and its status is COMPLETE, this repository is an incomplete upload. Shards appear incrementally; do not assume an early snapshot contains all results. COMPLETE means the publication inventory was traversed and all included artifacts were uploaded, with exclusions explicitly recorded. It does not mean that every historical experiment succeeded or every source file is included.

Download and restore

pip install huggingface_hub
hf download Seeker38/hibiki_temp --repo-type dataset --local-dir ./hibiki_temp
python3 ./hibiki_temp/restore_results.py \
  --repo-dir ./hibiki_temp \
  --destination /mnt/data/hieutm2/projects/hibiki_improve

Use a new destination if there is already a project there. The restorer refuses to overwrite different existing bytes and checks SHA256 of every restored file and chunk. No inference, ASR, training, metric recomputation or Slurm submission occurs during restoration. For a reproducible transfer, pin --revision to the completed repository commit shown in the source-side PUBLICATION_RESULT.json or the chosen Hub snapshot. Public downloads do not need the owner's upload token.

The data is stored as deduplicated 64-MiB-or-smaller chunks in roughly 512-MiB ZIP shards, rather than one enormous backup ZIP. Identical content is stored once. Downloading also needs room for the restored files; the downloaded shards and restoration coexist unless you remove the shards later yourself. Neither the uploader nor restorer deletes scientific originals.

Repository files

File	Meaning
DATASET_STATUS.json	Actual publication status, counts, CPU receipt and omissions count
FILE_INDEX.jsonl	Every included file's original relative/restore path, byte size, SHA256 and ordered chunk list
OBJECT_INDEX.jsonl	Chunk SHA256, size and containing shard
SHARD_INDEX.json	ZIP shard hashes, remote verification and commits
METADATA_SHA256.json	Checksums of publication metadata files
FOLDER_MAP.csv	Included file counts/bytes by top-level restoration directory
PATH_GUIDE.md	Interpretation of major experiment folders and original paths
OMISSION_INDEX.jsonl	Every examined excluded record and the reason; no secret values
ARCHIVE_CONTAINER_INDEX.json	ZIP containers whose unpacked members were preserved; duplicate ZIP wrappers themselves are not copied
SYMLINK_INDEX.json	Original project links, recorded but not automatically recreated across servers
PUBLIC_ASSETS_TO_DOWNLOAD.json	Exact public model/data/source pins for recovery; pretrained weights are not bundled from their external caches
ENVIRONMENT_PACKAGES.json, requirements-runtime.txt, environment-conda.yml	Recorded environment and bootstrap recipe, not a portable cloned Conda installation
shards/part-*.zip	Deduplicated file content; use the restorer rather than unpacking objects into a project

Coverage

Included eligible files cover completed and unsuccessful historical runs: executed code, protocols/amendments, configurations, token arrays, owners/clocks, saved outputs with identified generated-output locations, transcripts, metrics, figures, receipts, checkpoints and failure evidence. Original result values and execution statuses are preserved. A saved result or successful upload is not scientific validation.

ZIP task/return packages are inspected and their eligible members retained under _uploaded_archive_members/<original ZIP path>/<member path>. This keeps otherwise ZIP-only evidence while avoiding multiple copies of return wrappers. ZIP member content hashes are preserved; recompressing a wrapper will not reproduce its original ZIP hash. Original ZIP wrappers remain on the source server.

Excluded categories are explicit in OMISSION_INDEX.jsonl: redundant migration backups, Git/compiler/Python caches, authentication material or credential-pattern matches, private listener/identity records, original-source media or audio with unresolved provenance, opaque external corpus archives, invalid compressed files, and files that changed after inventory. This public repository is not a complete backup of those categories. Raw external datasets are not silently substituted with generated/synthetic input.

Current continuation state

The recent branch is C2R1/R4/C3N/M1. C2R1 has 16 historical policy rows plus four same-allocation H100 controls; historical MIG equality was not satisfied. R4's six saved-output speech assessments are completed. M1 has a completed CPU decision pack, but human listening remains NOT_REVIEWED; fresh originals/full human references/ancestry remain on hold; the native stop contract needs a versioned amendment. C3N generation remains disabled. Read the included final decisions and receipts rather than assuming an old pre-submission status file is the latest state.

All other experiment folders are historical evidence. A fresh server must recreate paths, public assets, compatible environment and device qualification before any separately authorized future run. Old Slurm launchers and consumed attempt markers are not instructions to submit again. See PATH_GUIDE.md and RIGHTS_AND_OMISSIONS.md.