米斯特拉尔人工智能测验

当您参加测验时，就会出现答案和解释。

1. A team selects a Mistral model for a hosted application. Which check should come first?
   • Assume every model in the family has the same capabilities
   • Inspect that exact model’s documented inputs, outputs, tools, limits, and intended use
   • Use a family name as a substitute for a version
   • Check only how many parameters it has
2. A downloadable model has open weights. Which conclusion follows?
   • Every use is automatically unrestricted
   • The model runs without compatible hardware or software
   • All models from the provider share one license
   • The parameters can be accessed, but license terms and deployment requirements still need review
3. Why might hosted API access and running weights locally require different reviews?
   • They involve different infrastructure, service terms, security and maintenance responsibilities
   • Hosted services never process data
   • Local systems have no security risks
   • The model’s version is irrelevant
4. A model has a large advertised context capacity. What does that establish about a long document task?
   • That every passage will be used correctly
   • That citations will always be accurate
   • Only a nominal capacity; the team should test reliable use of the expected material
   • That the model has no output limit
5. A quantized derivative is evaluated for a production workflow. Which record helps reproduce the result?
   • Only the provider’s family name
   • The exact derivative’s relationship to its base release, version and settings
   • A screenshot of the chat window
   • The base model card without the derivative details
6. A model version is being retired. What should a team do before relying on its replacement?
   • Review migration notices and retest representative tasks and failure cases
   • Assume the replacement behaves identically
   • Keep the old model name in logs and skip testing
   • Compare only marketing benchmarks
7. Two Mistral releases have different license terms. Which inference is justified?
   • The model with more features is always less restricted
   • Hosted use has no terms
   • Terms must be checked per release rather than generalized to the whole catalog
   • A downloadable release is always open for commercial use
8. A model passes a formatting test but gives weak answers in another language. What should selection depend on?
   • The provider’s broad reputation alone
   • The shortest model name
   • Whether the first prompt looked polished
   • Measured performance on the application’s representative tasks and requirements