810-110 sample questions with answers

1. A.

   Compare the providers' published per-token list prices only, because token price is the single cost that differs between hosting options and every other cost is identical.

2. B.

   Compare total cost of ownership at the expected steady utilisation, including hardware, power, networking and operations staff for local hosting, against the per-token charges for the same volume and required quality.

3. C.

   Compare the leaderboard scores of the cloud model and the open-weight model, choosing the higher-scoring one, because the most capable model always produces the lowest total cost.

4. D.

   Compare the purchase price of one GPU server against a single month of cloud API charges, choosing local hosting if the server costs less than two months of cloud usage.

1. A.

   Local hosting lets the insurer receive the cloud provider's updates earlier than other customers, so it can validate each new version before the provider releases it publicly.

2. B.

   The insurer holds the exact weights it validated, so the model changes only when the insurer chooses to upgrade and has re-validated, rather than on a provider's schedule.

3. C.

   Local hosting guarantees the model's outputs will never vary between runs, so the insurer can skip re-validation even after it decides to upgrade to newer weights.

4. D.

   Local hosting removes the need for validation altogether, because open-weight models are certified by their publishers for use in regulated claims workflows.

1. A.

   Neither model should be chosen, because a small model scoring the same as a large one proves that the evaluation set is broken and the results must be ignored.

2. B.

   The larger general model is the better fit, because larger models always deliver better results in production even when an evaluation on the team's actual security tasks shows equal scores.

3. C.

   The smaller domain-specific model is the better fit, because it matches the larger model on the tasks that matter at a fraction of the infrastructure cost and can run locally.

4. D.

   The smaller model should be chosen only if it is then fine-tuned to become as large as the general model, because domain models cannot be deployed at their original size.

Choose 2.

1. A.

   The number of times the model has been downloaded from the hub during the last month.

2. B.

   The intended uses and known limitations, including any languages or tasks the model is not suited to.

3. C.

   The number of community likes the repository has received since it was first published.

4. D.

   The evaluation results and the conditions they were measured under, including languages and task types.

5. E.

   The file names of the weight shards and the order in which they are loaded into memory.

1. A.

   Increase the model's temperature, so that it becomes creative enough to work with inputs that are longer than its configured context length allows.

2. B.

   Reload the model with a larger context length that the model supports, since shortening the input is not an option when every log line matters.

3. C.

   Ask the question a second time without changing anything, because context-length errors are transient and usually clear on the next attempt.

4. D.

   Switch the output format to bullet points, because shorter output makes room for the extra input tokens that exceeded the configured limit.

Choose 2.

1. A.

   Sending every prompt to the cloud provider for safety review before the local model answers it.

2. B.

   Paying the cloud provider's per-token charges for all traffic handled by the self-hosted model.

3. C.

   Monitoring the models' usage and behaviour, and applying patches and updates to the serving stack.

4. D.

   Accepting whatever model version the cloud provider deploys on its own release schedule.

5. E.

   Hardening and restricting access to the inference endpoints so outsiders cannot use or abuse them.

1. A.

   It starts from random noise and refines the whole image over a series of denoising steps, with the text prompt steering each step toward the requested content.

2. B.

   It classifies the prompt into one of a fixed set of image categories and returns the template image the model learned for that particular category.

3. C.

   It predicts the next pixel from all the pixels already placed, scanning row by row until the image is complete, exactly as a text model predicts tokens.

4. D.

   It retrieves the stored training image whose caption is closest to the prompt and then adjusts that stored image's colours, proportions and details until it matches the request.

1. A.

   Speech synthesis with an audio model, which narrates the room layout for accessibility so the ladder no longer needs to appear in the published version.

2. B.

   Image classification with a vision model, which labels the photograph as containing a ladder so that the intranet team can decide whether to publish it or not.

3. C.

   Text summarisation with a large language model, which describes the photograph without the ladder so that the description can be published in place of the image.

4. D.

   Image inpainting with a diffusion model, which regenerates a masked region so that it is consistent with the rest of the image.

1. A.

   Generate the report in smaller pieces, such as one section per request, and assemble the pieces, because a single response cannot exceed the model's maximum output length.

2. B.

   Lower the temperature to zero, because at zero temperature the model stops generating at random points less often and will then complete the long reports in full.

3. C.

   Move to a model with a larger context window, because the context window size is the only limit that decides how long a single generated response is allowed to become.

4. D.

   Raise the request's output-length setting further, because the model will honour any value the application supplies even when it is above the model's documented maximum for a single response.

1. A.

   Fixed-size chunking splits related log lines across chunk boundaries, so a retrieved chunk can hold the error without the preceding event that explains it.

2. B.

   The embedding model cannot process log files, because embeddings only work on well-formed prose and ignore any line that contains timestamps or error codes.

3. C.

   The language model deliberately omits triggering events, because summarising an incident requires reporting only the final error that the operator saw.

4. D.

   The vector database sorts chunks alphabetically, so the chunk holding the triggering event is always stored after the chunk that holds the error message.