Oratis's picture
Archive retired training source and historical synthetic research records
038efa5 verified
|
Raw History Blame Contribute Delete
2.78 kB
# Retired training source and research records
This owner-authorized supplement preserves **637 unique files / 15,393,470 bytes** of project-authored training/evaluation code, configuration, synthetic test labels and historical experimental records recovered from retired compute backups. It combines twelve fully audited home-directory archives and two nested source bundles. Another large home archive remains under review and is **not** claimed to be covered here.
`files.tar.gz` contains the files listed in `MANIFEST.jsonl`. Paths use the original SHA-256 plus the original basename so distinct historical versions coexist. `source_relative_paths` records the original project-relative locations with private home prefixes removed. Duplicate files already preserved in earlier public releases are not recopied into this supplement. There is no claim that this directory alone reconstructs a complete runnable checkout; see the [public source snapshot](https://github.com/CuddlerAI/Diogenes-Archive).
## Preservation and explicit changes
613 files preserve their exact original bytes. 24 documented privacy copies replace 420 private home-path occurrences, 387 internal-hostname occurrences and two private-network addresses. `PUBLICATION-COPIES.jsonl` lists original and public hashes, lengths and replacement counts. Original private evidence is retained separately. Across 6,268 valid JSON/JSONL records, key order, data types and all non-text values were checked unchanged. Historical numbers, losses, failures, training/evaluation distinctions and flags were not recomputed or improved.
Six files with a `.json` suffix originally contained JSON followed by prose. They remain historical text; `FORMAT_NOTES.json` identifies them. Consumers must not assume every `.json` filename contains a valid standalone JSON document.
Synthetic smoke-test labels and results describe pipeline checks and experimental runs, not independently validated real-world model capability. Upstream model names, evaluation-only flags, source attribution and recorded revisions remain as originally captured. A missing historical revision remains unknown.
## Scope and use
Downloaded base models, third-party datasets, Python environments, dependency caches, credentials, shell/operations state and private account information are excluded. Small self-trained model artifacts are tracked separately in the research-model archive; this text package contains no model weights. Private cloud archive contents are not made public wholesale.
Read `NOTICE.md` and the repository-level notice before reuse. `ARCHIVE_CONTENTS.json` gives the normalized archive SHA-256 and every member hash. Archive ownership and timestamps were normalized; file payloads match the published manifest exactly.