watsonx Data Scientist - Associate flashcards

32 cards. Read the term and the definition here. The deck at the top of this page is optional.

watsonx.data lakehouse

An open architecture that stores data in object storage in open formats such as Parquet and Iceberg, and queries it with Presto and Spark engines without copying into a proprietary warehouse.

Iceberg table

A table format on object storage providing ACID transactions, snapshots, schema evolution, and hidden partitioning, so multiple engines read consistent data concurrently.

Model factsheet

A governance document attached to a model asset that records purpose, intended use, training data, performance metrics, and known limitations for auditors and approvers.

Use case in governance

The audit anchor that groups related models and data assets under one business purpose, so approvals, factsheets, and evaluations are reviewed together.

Demographic parity

A fairness metric requiring the positive-prediction rate to be equal across protected groups; it compares selection rates, not true positive rates.

Equal opportunity

A fairness metric requiring equal true positive rates across groups, meaning qualified individuals are identified at the same rate regardless of group membership.

Concept drift

A change in the relationship between input features and the target over time, so a model that was accurate becomes inaccurate even if inputs look similar.

Local explanation

An explanation of one specific prediction, typically via SHAP values, showing which features pushed that individual record toward or away from the predicted outcome.

Prompt tuning

A tuning method that learns a small set of soft prompt vectors while freezing model weights, producing a compact artifact served with a prompt tuning id.

Fine-tuning

Updating all or adapter weights of a foundation model on labeled task data, requiring more data and compute than prompt tuning but allowing deeper behavior change.

Online deployment

A real-time scoring endpoint in a deployment space that autoscales with request volume and returns predictions synchronously per request.

Batch deployment

A deployment job that scores an entire input data file on a schedule or on demand, writing predictions to an output file rather than answering single requests.

Top-p sampling

Nucleus sampling that keeps the smallest set of tokens whose cumulative probability reaches p, then samples from that set; lower p makes output more focused.

Temperature

A decoding parameter scaling the randomness of token selection; near zero yields deterministic, repetitive output, while higher values increase diversity and risk.

Stop sequence

A string that terminates generation when produced, used to end structured output blocks and prevent the model from continuing past the intended answer.

Context window

The maximum number of tokens a model can process in one request, covering both the prompt and the generated output; exceeding it causes an error.

Few-shot prompting

Including several worked examples of input and expected output in the prompt so the model infers the task pattern and output format without weight updates.

Prompt variable

A placeholder such as {input} declared in a saved prompt asset and filled at run time, letting one template serve many inputs in notebooks or deployments.

AutoAI pipeline leaderboard

A ranked list of candidate pipelines scored on the chosen optimized metric, letting users compare preprocessing, algorithm, and hyperparameter combinations before selecting one.

AutoAI optimized metric

The single metric used to rank pipelines during an experiment, such as ROC AUC or root mean squared error; it must match the prediction type and business goal.

Watson Studio pipeline

A directed acyclic graph of nodes for data preparation, training, evaluation, and deployment that can be scheduled, making retraining reproducible and automated.

Retrieval-augmented generation

A pattern that retrieves relevant document chunks from a vector store and inserts them into the prompt so the model answers from supplied evidence rather than only its weights.

Chunk overlap

A setting that repeats a portion of text between adjacent chunks so sentences split at a boundary remain retrievable and context is not lost.

Embedding model consistency

The rule that the same embedding model and dimensionality must be used at index time and query time, otherwise similarity scores are meaningless.

Groundedness

A generation metric measuring whether the answer is supported by the retrieved context, distinguishing faithful responses from hallucinated content.

Agent tool selection

The model choosing among available tools from their names and descriptions, calling one, reading the result, and iterating until a final answer is produced.

Project versus space

A project holds authoring assets such as notebooks and prompts; a deployment space holds deployable assets and is the boundary for serving endpoints.

IAM bearer token

A short-lived access token obtained by exchanging an API key with the identity service, sent as an authorization header on watsonx API calls.

ModelInference class

The watsonx SDK class scoped by model_id and project_id or space_id that provides generate_text, chat, and embed_text methods for foundation models.

generate_text return shape

A dictionary containing generated_text plus input and generated token counts, unlike chat which returns choices with message content.

params dictionary

The SDK argument carrying decoding settings such as decoding_method, max_new_tokens, temperature, top_p, and stop_sequences; unknown keys raise validation errors.

Feedback dataset for monitoring

Labeled production data joined to predictions so quality metrics such as accuracy can be computed; without it only drift and input fairness are measurable.