No AI-specific product benchmark yet
The honest result is an open test plan
Current Dits measurements cover generic hashing and chunking components. They do not establish checkpoint reuse, dataset savings, experiment reproducibility, transfer cost, or training integration.
Modeled scenarios removed
Previous percentage-savings scenarios were not measurements and should not guide a storage purchase or architecture decision.
Required public workloads
- Consecutive checkpoints with diffuse weight updates.
- Base models with adapters, fine-tunes, and quantized variants.
- Append-heavy and mutation-heavy dataset snapshots.
- Interrupted ingest, corruption, checkout, and recovery.
- Storage, wall time, CPU, memory, read amplification, and exact output hashes.
- Equivalent runs with Xet, DVC, Git LFS, and plain object storage.