Prueba de pesas abierta

Las respuestas y explicaciones aparecen a medida que realiza el cuestionario.

1. ¿Qué distingue un modelo de 'pesos abiertos' de un modelo de código totalmente abierto según la definición OSI?
   • Se publican las ponderaciones, pero es posible que los datos, el código o los términos abiertos no se publiquen.
   • No hay diferencia significativa entre los dos términos.
   • Nadie puede descargar ni ejecutar los modelos de pesos abiertos
   • Los modelos de código abierto nunca liberan sus pesos entrenados
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. ¿Cuál es un beneficio importante de las pesas abiertas para los investigadores?
   • Pueden inspeccionar activaciones y ejecutar experimentos de ajuste.
   • Reciben puntuaciones más altas garantizadas en cada punto de referencia.
   • Ya no necesitan ningún hardware informático.
   • Obtienen acceso gratuito a los datos de entrenamiento de modelos cerrados
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