ralomi-reazonspeech-20h-development

Ralomi development checkpoint backup

This repository is configured as public with manual access approval. No audio, transcripts, or training manifests have been uploaded yet. The original transfer-staging metadata is retained below as history. Since 2026-10-08 this repository also preserves encrypted saved fullstate checkpoints from the actual 5000h development training and its continuations.

Source and terms

Source: ReazonSpeech, provided by Reazon Human Interaction Laboratory / Reazon Holdings, Inc. Pinned upstream revision: 0df78f991f6afa051aacb7c63e3fe2b496c13c2c. License: Community Data License Agreement – Sharing, Version 1.0. Upstream specifies use solely within Article 30-4 of the Japanese Copyright Act; review the upstream dataset card and comply with applicable law. The license is not altered by this access gate. No additional license restrictions are asserted here.

Planned data identity

Prepared train: 13,184 items / 20 hours. Prepared validation: 1,319 items / 2 hours. The existing 662-item evaluation exclusion and train/validation separation are retained. No full source corpus, unrelated evaluation audio, credentials, environment files, or teacher outputs are included. A future data upload must retain attribution and document filtering, normalization and tokenization changes, checksums and the exact file inventory. The original staging repository contained metadata only. The model-backup files added on 2026-10-08 contain encrypted checkpoints, not corpus data.

Periodic encrypted model backups

Saved production checkpoints are checked every five minutes by the owner workstation while powered on and logged in. Completed files older than 60 seconds are eligible. Model, optimizer, scaler and distributed RNG fullstates are preserved. No audio, transcripts, manifests, credentials, private logs or decryption keys are uploaded. No access requests are automatically approved. This repository is archival storage; no model release, deployment qualification or downstream rights are asserted.

Archives under encrypted/models/ use OpenSSL AES-256-CBC with salted PBKDF2 and 200000 iterations. The key is held only by the owner. Each encrypted tar contains inventory.json with the original checkpoint paths, steps and SHA256. Restore with openssl enc -d -aes-256-cbc -pbkdf2 -iter 200000 -pass file:LOCAL_KEY -in ARCHIVE -out RESTORED.tar, then verify each inventory member SHA256 before use.

Backups are accepted only after native strict finite loading, source/member SHA verification, pinned Hub ciphertext readback, decrypted tar/member SHA verification, manual gating and anonymous binary denial. The current source/trainer remain intact; saving a checkpoint does not rerun training or evaluation. Network failures retry at most three times in a finite cycle and the next scheduled cycle catches up with outstanding checkpoints. Manual sleep or power-off suspends workstation-based uploading until it resumes.