MVE research checkpoint archive
Public archival of the owner's MVE experiment artifacts. This is an evolving research archive, not a claim of benchmark leadership or a standard Transformers inference package.
migration/20261019/manifest.json maps original run identities and training-state / exact inference variants to immutable SHA-256 checkpoint objects. Each migrated checkpoint is fully uploaded and streamed back at its recorded commit to verify every byte before its redundant local F-drive copy is removed. Full training checkpoints retain optimizer and RNG state; inference capsules are identified separately and do not imply training resumability. Protocols and model families remain independently frozen.
provenance/<sha256>.json records original verification receipts and source identities. server-manifests/ contains directly archived MVE2 finals. No credentials, original datasets or feature caches are included. Scientific scores require their corresponding strict evaluation audit; incomplete training-seed sets are not averaged.
The archive preserves original bytes. Load trusted PyTorch checkpoints only in a compatible environment. License and attribution of the original pretrained components continue to apply; this archive does not grant additional rights to third-party assets.