Fungua Maswali ya Uzito

Majibu na maelezo huonekana unapojibu maswali.

1. What distinguishes an 'open-weights' model from a fully open-source model under the OSI definition?
   • Weights are released, but data, code or open terms may not be
   • There is no meaningful difference between the two terms
   • Open-weights models cannot be downloaded or run by anyone at all
   • Open-source models never release their trained weights
2. A team can download a model checkpoint. What should it check before using the model in a product?
   • Assume downloading grants every commercial and redistribution right
   • The permissions and restrictions in the license for that exact release
   • Rely only on the model family name
   • Treat the download page as a replacement for license terms
3. Why might a downloaded model need more than its weight file for compatible inference?
   • A weight file must include every training example
   • Only the original author can load any model
   • Architecture, tokenizer or preprocessing, configuration and runtime determine how weights are interpreted
   • A model cannot run without a cloud account
4. How can fine-tuning affect a model’s refusal behavior?
   • It guarantees every refusal is preserved
   • It makes the model permanently unable to answer benign requests
   • It remotely updates the original provider’s hosted model
   • It can change refusal behavior, so the adapted model should be reevaluated
5. What is an important benefit of open weights for researchers?
   • They can inspect activations and run fine-tuning experiments
   • They receive guaranteed higher scores on every benchmark
   • They no longer need any computing hardware at all
   • They get free access to closed models' training data
6. What should an operator keep when modifying an open-weight model?
   • Only a public download counter
   • A record of modifications and a rollback version
   • A claim that all adaptations are compliant
   • The base model’s marketing slogan
7. Which claim is supported by access to model weights alone?
   • The training data are public
   • The model can be reproduced exactly
   • The parameters can be downloaded under the release’s terms, but other components may still be missing
   • Every use is permitted commercially
8. What should a team check before using a quantized open-weight checkpoint?
   • Only the model family name
   • That quantization guarantees identical quality
   • That the checkpoint needs no tokenizer or runtime
   • Its task behavior and resource use on the intended hardware