Model Serialization Formats Quiz

Answers and explanations appear as you take the quiz.

1. Why is Python's pickle format considered a security risk for models?
   • Loading it can execute arbitrary code embedded in the file
   • It compresses files too much
   • It only works on weekends
   • It deletes the original model
2. What is the main advantage of the ONNX format?
   • It is the smallest possible format
   • It lets a model trained in one framework run in many other runtimes and languages
   • It can only be opened in PyTorch
   • It automatically improves accuracy
3. Why has safetensors become popular for sharing model weights?
   • It runs code on load for convenience
   • It is the oldest format available
   • It stores only tensor data safely, with fast memory-mappable loading and no code execution
   • It only stores model accuracy
4. GGUF is most associated with which use case?
   • Storing spreadsheets
   • Encrypting emails
   • Training models from scratch in the cloud
   • Running quantized large language models efficiently on local hardware
5. What does an interchange format like ONNX primarily store?
   • A standardized computational graph and the model's weights
   • A video of the training process
   • Only the training dataset
   • The developer's source code comments
6. When exporting a Keras 3 model for TensorFlow serving, which directory format can model.export produce?
   • GGUF
   • SavedModel
   • CSV
   • A pickle file containing the entire Python application
7. After exporting a model to ONNX for a new runtime, what should the team check?
   • That the ONNX file guarantees identical speed on every device
   • That the export automatically includes the original training data
   • That supported operators run and outputs match the reference model on representative inputs
   • That no preprocessing code is needed
8. Why can a state-dictionary checkpoint be preferable to loading a whole Python object from an untrusted file?
   • It guarantees the model architecture is embedded
   • It makes every checkpoint safe regardless of loader and version
   • It automatically converts the model to ONNX
   • It can load tensor state without reconstructing arbitrary pickled Python objects, subject to the loader’s security settings