Prueba de ajuste

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

1. ¿Qué cambia durante el ajuste, a diferencia de agregar ejemplos a un mensaje?
   • Learned parameters of an existing model are updated through further training
   • Only the current request text changes, while model parameters stay fixed
   • A current document is fetched at answer time without changing the model
   • The model's output is edited after generation with no training
2. Un asistente de soporte escribe claramente pero debe utilizar una política actualizada cada semana. ¿Qué primera prueba se ajusta mejor a la guía?
   • Retrain the model after every policy edit before considering other options
   • Put an old policy into training examples and rely on model recall
   • Retrieve the current policy at answer time and evaluate its use
   • Tune only the model's output style and assume policy facts remain current
3. Un modelo omite repetidamente un formato de respuesta estable. ¿Qué comparación debería preceder a una decisión de ajuste?
   • Only compare training loss across fine-tuned checkpoints
   • Compare prompt changes with a carefully evaluated fine-tuning dataset
   • Replace the held-out evaluation set with more training examples
   • Assume retrieval of unrelated documents will fix formatting
4. ¿Por qué el conjunto de evaluación debería permanecer separado tanto de los datos de ajuste como de las decisiones de ajuste?
   • To make the training set contain fewer labels by definition
   • To guarantee that every output will cite a source
   • To avoid recording the base model version
   • To measure behavior on examples that were not used to fit or select the model
5. ¿Qué comprobaciones de conjuntos de datos requiere la guía antes de realizar ajustes?
   • Only whether the training job can parse every record
   • Labels, duplicate records, permissions and confidential information
   • Only whether examples have the desired output length
   • Only whether the dataset is larger than the model's context window
6. A method trains only a subset of model parameters. What requirement still applies?
   • Evaluate the target task and capabilities that should remain intact
   • Assume lower memory use proves equal quality on every task
   • Skip the held-out set because fewer weights changed
   • Treat training loss as a complete safety evaluation
7. The fine-tuned model improves the target classifier but performs worse on unrelated tasks. What does that show?
   • That the model must have received an invalid prompt
   • That held-out tests are unnecessary if training loss falls
   • That retrieval can never work with a fine-tuned model
   • A gain in one measured behavior can come with a regression elsewhere
8. What should a team record to make a fine-tuned release reviewable and reversible?
   • Only the final training loss and public model name
   • Only the prompt used during one successful demonstration
   • Base model, dataset version, training settings, checkpoint and rollback plan
   • Only the number of examples and the selected learning rate