Training output delivery
How do you send AI models, checkpoints, and datasets?
A model often depends on more than its largest weight file. Configuration, tokenizer data, shard indexes, checksums, and run instructions need to travel with it. Keep those files together when handing work to a teammate.
Organize model files, configuration, tokenizer data, examples, checksums, and documentation into one folder. Remove data and credentials that should not be shared, then send the file or complete folder. The receiver should verify it before loading the model.
What belongs in the model folder?
- Weights or sharded checkpoints and their index files.
- Model configuration, tokenizer data, preprocessing settings, and dependency versions.
- Checksums, license terms, data restrictions, and source notes.
- A minimal loading example and the required training or quantization details.
Review sensitive content before transfer
Check whether training data contains personal information, trade secrets, or restricted material, and confirm that you can send both the model and dataset. Remove API keys, access tokens, experiment tracking credentials, and local caches. Peer-to-peer transfer does not replace your access review.
Send large weights and shard folders together
Desktop Chrome and Edge support a single file up to 200 GB and folders containing up to 20,000 files. This works for a live handoff of safetensors files, checkpoint shards, configuration, and test examples to a teammate who is online now.
When should you use a model registry or object storage?
Use an approved model registry, data platform, or object store for versioning, access control, persistent downloads, multi-user work, or automated deployment. 8ma Quick Transfer handles a one-time live delivery; it does not replace a model registry.