Datasets used for training must be reproducible. What should you register?
Why B is the answer
Versioned data assets
Versioned data assets record exactly what a model was trained on.
In detail
- A versioned data asset identifies exactly which data was used.
- Record the version with the model.
Use case
An illustrative scenario, not a real incident.
A model cannot be reproduced, because its data folder was overwritten. A versioned data asset names exactly which data was used.
See it in different ways
mindmap
root((Reproducible run))
Code
Git
Data
Versioned data assets
Environment
Registered environment
Compute
Cluster that scales to zero
Results
MLflowstateDiagram-v2 state "Data version 1" as s0 state "Data version 2" as s1 state "Training job" as s2 [*] --> s0 s0 --> s1: new data s1 --> s2: reference by version s2 --> [*]
Key point: Version data as well as code.
Why the other options are wrong
- A. The latest folder, overwritten on each run
- Overwriting a folder loses the old data.
- C. A datastore pointing at the raw folder
- A datastore connects storage, and it does not fix a version.
- D. A list of file names in the job's description
- A list in a description is not a data version.