AI-300 · Domain 5: Optimize generative AI systems and model performance · 13% of the exam
Task 5.2: Implement advanced fine-tuning and model customization
Advanced fine-tuning methods, synthetic data for fine-tuning, monitoring and optimizing fine-tuned models, and taking a fine-tuned model from development to production deployment.
Study it
Fine-tuning methods: supervised, preference and reinforcement
Try each one before opening the answer. Every option is explained, with the Microsoft documentation page that proves it.
Question 1 · choose 1
A bank fine-tunes a model in its West Europe Foundry resource. Regulators require training data and weights to stay in the resource's region, cost matters, and jobs may take a little longer. Which training type fits?
AGlobal training, which is cheaper and moves the data only for the duration of the job
BDeveloper training, which uses idle capacity and is the cheapest option of all
CStandard training, which runs in the resource's region with data residency guarantees
DGlobal training with a private endpoint on the resource to keep the data in the region
Show the answer and why
AGlobal training, which is cheaper and moves the data only for the duration of the job
Incorrect
Global copies data and weights to the region where training occurs.
BDeveloper training, which uses idle capacity and is the cheapest option of all
Incorrect
Developer training is preemptible and has no data residency guarantee.
CStandard training, which runs in the resource's region with data residency guarantees
Correct
Only Standard keeps training in the region with residency guarantees.
DGlobal training with a private endpoint on the resource to keep the data in the region
Incorrect
Private networking doesn't change where Global training runs.
Fine-tuning training types decide where training runs. Standard training occurs in the current Foundry resource's region with data residency guarantees. Global training is more affordable by using capacity beyond the region, copying data and weights to where training occurs. Developer training uses idle capacity with no latency, SLA or residency guarantees and can be preempted.
After an SFT job, train_loss kept falling every epoch, while full_valid_loss rose after epoch 2 of 4. The final model looks worse on held-out tickets. What should the team do?
ADeploy the final model; the falling training loss shows that it learned the most
BCompare the checkpoints on validation metrics and held-out tasks, and deploy the earlier, better one
CDouble the number of epochs so that the validation loss and held-out results have time to recover
DRemove the validation file so that full_valid_loss stops being reported
Show the answer and why
ADeploy the final model; the falling training loss shows that it learned the most
Incorrect
Training loss alone doesn't show quality on new data.
BCompare the checkpoints on validation metrics and held-out tasks, and deploy the earlier, better one
Correct
Rising validation loss with falling training loss signals overfitting, so an earlier checkpoint can be better.
CDouble the number of epochs so that the validation loss and held-out results have time to recover
Incorrect
More passes tend to increase overfitting.
DRemove the validation file so that full_valid_loss stops being reported
Incorrect
Removing validation hides the problem instead of fixing it.
SFT reports train_loss, full_valid_loss and token accuracy metrics. Compare training and validation results rather than training performance alone: if training loss falls while validation loss rises, inspect failing validation examples. More epochs increase overfitting risk. Compare the available checkpoints using validation metrics and held-out task results before choosing one to deploy.
A team prepares preference data for direct preference optimization. What should each JSONL example contain?
AA prompt, a reference answer and a numeric reward from 0 to 1 for each
BA conversation with one assistant answer and a quality label for it
CTwo different prompts and one assistant answer that fits both of them
DAn input messages array with a preferred_output and a non_preferred_output
Show the answer and why
AA prompt, a reference answer and a numeric reward from 0 to 1 for each
Incorrect
Rewards from graders belong to reinforcement fine-tuning.
BA conversation with one assistant answer and a quality label for it
Incorrect
That's closer to labeled SFT data, not preference pairs.
CTwo different prompts and one assistant answer that fits both of them
Incorrect
Both outputs must answer the same input.
DAn input messages array with a preferred_output and a non_preferred_output
Correct
DPO examples pair a preferred and a rejected response.
For DPO, save training.jsonl and validation.jsonl with one example per line. Each example contains an input (a messages array with the prompt, plus optional tools), a preferred_output and a non_preferred_output, both answering the same prompt with at least one assistant message.
An SFT job trains on a file of 2 million tokens for 3 epochs. It waited 5 hours in the queue, and the training itself took 90 minutes. How is the training charge calculated?
A6.5 hours of elapsed time, the queue included, times the hourly training rate for the model
B6 million training tokens times the training price per token; queue time isn't charged
C90 minutes of training time times the hourly training rate
D2 million tokens times the deployed model's input token price
Show the answer and why
A6.5 hours of elapsed time, the queue included, times the hourly training rate for the model
Incorrect
Queue time isn't charged, and SFT isn't billed by time.
B6 million training tokens times the training price per token; queue time isn't charged
Correct
SFT is billed on training tokens times epochs, not on time.
C90 minutes of training time times the hourly training rate
Incorrect
Hourly training billing applies to reinforcement fine-tuning, not SFT.
D2 million tokens times the deployed model's input token price
Incorrect
Training token prices differ from inference prices, and epochs multiply the tokens.
Supervised fine-tuning and DPO use token-based billing: price = training tokens x epochs x training price per token. You aren't charged for time in queue, failed jobs, jobs canceled before training begins, or data safety checks, and the elapsed time doesn't determine an SFT charge. Training token prices differ from inference prices; RFT uses training-time billing instead.
In Foundry Managed Training, what does the team provide?
AA GPU cluster sized for the model
BThe training data and the job settings
CA Python training loop with custom losses
DA container image with the training code
Show the answer and why
AA GPU cluster sized for the model
Incorrect
Foundry-managed infrastructure runs the job.
BThe training data and the job settings
Correct
No training GPUs are provisioned by the team.
CA Python training loop with custom losses
Incorrect
That's Programmed Training.
DA container image with the training code
Incorrect
Managed Training doesn't need an image.
Managed Training runs a training job from your data and settings on Foundry-managed infrastructure, so you don't provision training GPUs. Programmed Training (preview) lets you control training steps in Python.
Run lifecycle commands only when the job supports them.
Run lifecycle commands such as pause, resume and cancel only when your model, method and job state support them; pause and resume aren't available for every model or method.
What does setting detail to low on training images do in vision fine-tuning?
ASkips images from training
BRaises the resolution so that the model sees fine detail
CDownscales images to 512x512 to cut tokens and cost
DConverts images to grayscale
Show the answer and why
ASkips images from training
Incorrect
Images are still used.
BRaises the resolution so that the model sees fine detail
Incorrect
That's high.
CDownscales images to 512x512 to cut tokens and cost
Correct
low reduces cost but may lose fine detail.
DConverts images to grayscale
Incorrect
detail controls resolution, not color.
The detail parameter in image_url controls image fidelity and training cost: low downscales images to 512x512 pixels and uses fewer tokens, while high processes images at full resolution with more tokens.
Which data zone does Data zone fine-tuning training currently support?
AOnly the US data zone
BThe EU and US data zones
CAny data zone the resource chooses
DNone; data zone training was retired
Show the answer and why
AOnly the US data zone
Correct
Data zone training currently supports the US zone only.
BThe EU and US data zones
Incorrect
Only the US zone is listed.
CAny data zone the resource chooses
Incorrect
Support is limited.
DNone; data zone training was retired
Incorrect
It's available for the US zone.
Data zone training occurs within a supported data zone rather than a single region, for workloads where data must stay within the zone. Only the US data zone is currently supported.