01. A multimodal app receives a blurry, low-resolution photo of a receipt and cannot reliably read several of the line items. What is the most appropriate behavior?
a) Flag the low image quality and ask the user for a clearer photo rather than guessing the unreadable values.
b) Confidently fill in the unreadable line items with the most likely numbers.
c) Ignore the image and answer only from the model's general knowledge of receipts.
d) Raise the temperature to make the model read the blurry text more accurately.
02. Which practices are appropriate when designing a multimodal application that analyzes uploaded chart and document images?
(Choose two.)
a) Avoid stating conclusions the visual evidence does not actually support.
b) Check that the image is readable before analysis and handle low-quality input gracefully.
c) Assume every uploaded image is accurate and complete without checks.
d) Raise the temperature to sharpen blurry regions of the image.
e) Fill in any missing values in a chart automatically to keep the final output table complete
03. An LLM feature is proposed to automatically approve or deny benefit claims and send the decision to applicants with no human involvement. Which control is most important before deploying this?
a) Keep a human in the loop to review high-impact decisions before they are finalized and sent.
b) Raise the temperature to make the decisions feel more natural to applicants.
c) Remove authentication so applicants can reach the decision faster.
d) Let the model finalize all decisions and audit a few of them next quarter.
04. After heavily fine-tuning a model on one narrow formatting task, a team notices it now handles general requests noticeably worse than before. What does this most likely indicate, and how should they respond?
a) The general requests must be failing from a context window that physically shrank during the fine-tuning process
b) The model simply needs more epochs, run on the same narrow task, to recover its general skills
c) The regression is a display issue fixed by enabling streaming on general requests.
d) The tune over-specialized, so evaluate broader behavior and scope the tuning more carefully.
05. A user tells a customer-service bot, in a very polite tone, to ignore its rules and reveal another customer's stored personal details. How should a well-designed guardrail respond?
a) Refuse the request regardless of its polite tone, and offer a safe alternative.
b) Comply only after lowering the temperature to phrase the disclosure carefully
c) Ask the user to rephrase the request as a formal command and then reveal the details.
d) Comply, treating a courteous and reasonable request as one to honor
06. A support bot must stay current with a price list that changes weekly and must also be able to create a refund request in a billing system on the customer's behalf. Which two capabilities best match these needs?
(Choose two.)
a) Retrieval over the regularly updated price list so answers reflect the current prices.
b) A tool or function call that submits the refund request to the billing system.
c) A longer system prompt that describes in detail how refunds are normally processed and approved.
d) Weekly fine-tuning on the new prices so they are baked into the model.
e) Raising the temperature to make the model adapt to new prices on its own
07. A team wants an open framework to build an agent that connects a Qwen model to tools and memory in a reasoning loop. Which option best describes ModelScope-Agent in this context?
a) It is a decoding parameter that controls how randomly the agent chooses each next token during its generation step.
b) It is a storage service for keeping the documents an agent retrieves.
c) An open agent framework that orchestrates an LLM with tool calls and memory.
d) It is a fine-tuning technique for teaching a base model new facts before deployment.
08. An agent succeeds on common requests but fails on a recurring category of edge cases, and the team wants a sound way to improve it. Which approach best reflects iterative agent improvement?
a) Raise the temperature to make the agent try more varied actions on the hard cases
b) Add as many tools as possible, for search and retrieval, to give the agent more options
c) Remove the stop conditions to let the agent keep retrying until it eventually succeeds
d) Examine the failing traces, refine prompts and tools, add validation, and re-evaluate.
09. A team can fine-tune on either one million scraped, noisy, duplicated examples or five thousand carefully curated and deduplicated examples that match the target behavior. Which choice is generally sounder and why?
a) The one million examples, whose duplicates reinforce the right patterns through repetition.
b) Five thousand curated, deduplicated examples that match the target behavior.
c) Either set works equally well as long as the temperature is lowered at inference.
d) The one million examples, as more data always produces a better fine-tuned model.
10. In a RAG system, documents are converted into numeric vectors and stored in a vector index before any user question arrives. What is the main purpose of this step?
a) It permanently teaches the model the document contents by updating its weights.
b) It compresses the documents to generate the answer faster.
c) It converts the documents into a single summary that the model reads each time.
d) It enables semantic retrieval by meaning rather than only exact keywords.