Train and customize ML and AI models

• Script mode runs custom training code inside a supported framework environment
• Training dependencies must be available in the remote framework environment
• Training scripts read data through the job's configured input channels
• Final model artifacts must be written to the designated model output directory
• Bayesian tuning uses previous results to choose promising configurations
• Random search supports independent concurrent tuning trials
• Grid search enumerates the supplied categorical combinations
• Hyperband reallocates training resources using intermediate trial performance
• A tuning job optimizes one selected objective metric in the specified direction
• Custom tuning metrics must be emitted and matched by metric definitions
• Logarithmic search is suitable for positive ranges spanning orders of magnitude
• Total trial limits and concurrent trial limits control different tuning resources
• Warm-start tuning reuses search results rather than retained training infrastructure
• Warm-start type determines whether the tuning dataset can change
• Managed warm pools reduce startup time between matching training jobs
• Warm-pool reuse requires matching resource and security settings
• The warm-pool persistent cache survives matching jobs but not pool termination
• Data parallelism distributes examples across replicas of a model
• Model parallelism partitions a model that does not fit on one device
• Early stopping saves training work when validation performance stops improving
• Managed Spot Training trades interruption tolerance for lower training cost
• Checkpoints let interrupted training resume instead of restarting from the beginning
• More training workers do not guarantee lower training time or cost
• Strong training performance with weak validation performance indicates overfitting
• Poor training fit calls for learning capacity or training evidence before stronger regularization
• Weight penalties and dropout reduce overfitting through different mechanisms
• Mixing general-capability examples into fine-tuning can reduce catastrophic forgetting
• Epochs count dataset passes while steps count training iterations
• Reducing batch size can resolve training out-of-memory failures
• Learning rate trades update speed against convergence stability
• Learning-rate warmup gradually introduces the target update size
• Bagging combines separately trained learners rather than a sequential correction chain
• Stacking learns how to combine base-model predictions
• Validation performance determines whether a base model or stacked ensemble is preferable
• Intelligent prompt routing selects a model per request rather than combining generated answers
• Few-shot prompting provides task examples without a training job
• A task prompt needs both the instruction and the material to operate on
• Temperature changes sampling variation rather than supplying missing knowledge
• Top K limits candidate count while Top P limits cumulative probability mass
• Output limits and stop sequences terminate generation for different reasons
• Direct model requests need the relevant conversation history in the current input
• Generation parameter support and ranges depend on the selected model
• Prompt templates separate reusable instructions from changing inputs and support controlled comparisons
• Hybrid retrieval combines raw-text matching with vector similarity
• Metadata filters constrain which source documents are eligible for retrieval
• Reranking reorders retrieved chunks using a relevance model
• The requested retrieval result count is a maximum rather than a guarantee
• Logical document segments can improve retrieval even when the embedding model accepts longer input
• Multilingual input support does not guarantee strong cross-language retrieval
• Bulk indexing and interactive query embedding have different performance priorities

Unlock with Premium — includes all practice exams and the complete study guide.