Implement text analysis solutions 4.1 Apply language model text analysis
A bank will run managed fine-tuning on customer-service transcripts.
Study guide skill: Customize language model outputs for domain tasks, such as compliance summarization and domain extraction
Question B7-011
A bank will run managed fine-tuning on customer-service transcripts. Its policy requires that training data stays in the Foundry resource's own region, and training cost and queue time are secondary. Which training type should it choose?
Select one answer.
Answer: C
Standard training runs in the Foundry resource's own region and guarantees data residency, so it suits data that must stay in one region. Global is cheaper and can have shorter queues, but it copies data and weights to whatever region runs the training. Developer uses idle capacity, can be preempted, and gives no residency guarantee, and Data zone keeps data only within a supported data zone, not a single region.
Deep dive
How it works
Managed fine-tuning in Foundry lets you choose a training type. It decides where training runs and therefore where your data goes. The four types trade cost and queue time against residency:
- Standard: training runs in the Foundry resource's own region, with data residency guarantees. Meant for workloads where data must stay in a specific region.
- Data zone: training runs within a supported data zone rather than a single region. Only the US data zone is currently supported.
- Global: cheaper than Standard because it uses capacity beyond your region, with faster queue times. Data and weights are copied to the region where training runs.
- Developer: big savings from idle capacity, but jobs can be preempted and resumed, with no latency or SLA guarantees and no data residency guarantee.
The bank's rule, "data stays in the resource's own region", is only satisfied by Standard. Cost and queue time are secondary, so Standard's higher price and possibly longer queues are acceptable.
Training type is separate from deployment type. Once the model is trained, choose a deployment type that also meets the residency rule.
The other options, and when you'd use them
- Global: Data and weights are copied to wherever training occurs. Use it when residency isn't a restriction and you want lower cost and faster queues.
- Data zone: It keeps data within a data zone, not a single region. Use it when zone-level residency is enough.
- Developer: There's no guarantee for data residency, and jobs can be preempted. It's for experimentation and price-sensitive work.
Exam tip
Fine-tuning training types: Standard = single-region residency; Data zone = zone residency; Global = cheapest and fastest, data moves; Developer = idle capacity, preemptible, no residency.
References
Last reviewed: