liveBook · Manning

Baseline

Also called reference point

A baseline is a reference model, result, measurement, or level used to compare performance or assess change.

How 27 Manning books explain it

Machine Learning System Design

A feature baseline is a baseline that makes predictions from a selected set of input features using a comparatively simple modeling approach.

How it is used
Uses baselines to connect ML system components, test product assumptions, and provide a reference for improvement.
How deep it goes
Explains baseline selection and comparison through practical tradeoffs and design examples rather than algorithm derivations.
Examples
Illustrates demand forecasting with lagged sales and image search with TF-IDF features and a linear relevance model.
What you can do after
Evaluate baseline candidates using accuracy, effort, interpretability, and computation time, and compare later models against their errors.

Read how Machine Learning System Design explains baseline →

Rearchitecting LLMs

A quantitative baseline is an objective reference measurement established before interventions to evaluate their effects.

How it is used
Measures model performance before optimization to assess intervention effects and trade-offs.
How deep it goes
Provides a procedural workflow for selecting metrics, validating a harness, recording measurements, and repeating tests under matching conditions.
Examples
Uses JSON schema compliance, a distilled student challenge, and a vLLM cache and latency probe as reference measurements.
What you can do after
Practice establishing task, general-capability, and efficiency baselines and comparing modified models with the same test battery.

Read how Rearchitecting LLMs explains baseline →

Delivering Machine Learning Projects: From design to deployment

A baseline performance measure is an initial quantitative assessment against which later model performance can be compared.

How it is used
Establishes an early performance reference for judging modelling potential, progress, and the value of more sophisticated approaches.
How deep it goes
Defines a baseline ticket and simple modelling choices, while deferring detailed performance estimation methods.
Examples
Uses majority-class accuracy and a quick regression for inventory demand and customer churn to illustrate baseline measures.
What you can do after
Evaluate whether later models improve on simple alternatives enough to meet business needs and justify their cost.

Read how Delivering Machine Learning Projects: From design to deployment explains baseline →

Deep Learning with R, Third Edition

A baseline is a reference level of performance produced by a simple strategy or comparison group against which a model's performance is evaluated.

How it is used
Establishes a reference score for judging early model progress and justifying additional complexity.
How deep it goes
Defines commonsense baselines and works through classification and forecasting comparisons, including an R implementation of forecast error.
Examples
Compares majority-class accuracy and a current-temperature forecast with learned models.
What you can do after
Evaluate model performance against simple reference strategies and reconsider the approach when reasonable models fail to beat them.

Read how Deep Learning with R, Third Edition explains baseline →

Machine Learning for Tabular Data: XGBoost, Deep Learning, and AI

A deliberately simple predictive model used to establish a reference level of performance.

How it is used
Establishes a performance measuring stick before optimizing an XGBoost price-prediction model.
How deep it goes
Explains benchmarking, overfitting checks, pipeline validation, cross-validation predictions, metrics, and coefficient inspection; defers complex-model optimization.
Examples
Uses transformed Airbnb Tokyo listings, linear-regression predictions, diagnostic plots, coefficient reports, and a metric comparison with optimized XGBoost.
What you can do after
Evaluate/understand whether a complex tabular model improves on a simple baseline and inspect its diagnostic failures.

Read how Machine Learning for Tabular Data: XGBoost, Deep Learning, and AI explains baseline →

Effective Data Analysis

A baseline is a reference measurement of a metric's current or historical performance before changes or interventions.

How it is used
Establishes a starting point for strategic decisions and metric goals before changes or interventions.
How deep it goes
Examines historical distributions, trends, anomalies, segments, time patterns, and correlations without prescribing one universal analysis.
Examples
Uses customer login activity by month and region, plus footwear sales compared with an external spending benchmark.
What you can do after
Evaluate/understand historical metric behavior to inform target achievability and useful time aggregation.

Read how Effective Data Analysis explains baseline →

Bayesian Optimization in Action

A reference point is a fixed objective-space point relative to which dominated hypervolume and hypervolume improvements are computed.

How it is used
Measures multiobjective optimization progress by anchoring dominated hypervolume at a chosen worst-case objective vector.
How deep it goes
Defines the lower-bound role, gives coordinate-selection guidance and numerical examples, then demonstrates BoTorch computation.
Examples
Uses [-2.0292, -0.4444] for two toy objectives and [-1.5, -2] for airplane design.
What you can do after
Apply it via BoTorch’s DominatedPartitioning to compute dominated hypervolume and compare optimization policies.

Read how Bayesian Optimization in Action explains baseline →

AI Recommender Systems

A popularity baseline recommends items according to their overall popularity and serves as a reference for personalized methods.

How it is used
Uses nonpersonalized top-item recommendations to test whether a recommender improves beyond general popularity.
How deep it goes
Works through training-set item counts and an offline precision comparison; calls for further ranking, coverage, and segment checks.
Examples
Compares ALS precision@10 of 0.1523 with popularity precision@10 of 0.1234 and supplies a pandas recommendation function.
What you can do after
Build it from scratch by counting training-set item interactions, ranking the top items, and evaluating identical per-user lists.

Read how AI Recommender Systems explains baseline →

Deep Learning with Python, Third Edition

A simple reference method or model used to judge whether a machine-learning system outperforms a naive or trivial alternative.

How it is used
Uses a simple reference as an early sanity check on model development and progress.
How deep it goes
Explains how to select and beat a baseline, with a worked forecasting evaluation and guidance for unsuccessful models.
Examples
Compares majority-class accuracy and a current-temperature forecast against learned models.
What you can do after
Evaluate whether a model improves on a task-appropriate trivial alternative and whether extra complexity is justified.

Read how Deep Learning with Python, Third Edition explains baseline →

LLM Customization and Fine-Tuning

A reference performance level against which model changes are evaluated.

How it is used
Uses measured performance to decide whether model adaptation is necessary and to compare subsequent variants.
How deep it goes
Provides a measurement procedure, sample-size guidance and decision paths, including uncertainty limits for small evaluation sets.
Examples
Uses support-ticket classification and a study-guide model comparison to illustrate baseline-driven decisions.
What you can do after
Evaluate whether a model change improves on a measured reference and meets a business requirement.

Read how LLM Customization and Fine-Tuning explains baseline →

Reinforcement Learning from Human Feedback

A baseline is a reference value subtracted from a reinforcement-learning return or reward estimate to reduce gradient-estimate variance without changing its expectation.

How it is used
Reduces noisy policy-gradient updates by expressing rewards or returns relative to a reference value.
How deep it goes
Derives action-independent gradient preservation and leave-one-out advantages, then implements the latter with reward tensors.
Examples
Uses multiple prompt completions and their rewards to demonstrate leave-one-out advantage calculation.
What you can do after
Evaluate baseline choices for variance reduction and recognize when action dependence affects gradient unbiasedness.

Read how Reinforcement Learning from Human Feedback explains baseline →

Generative AI in Action

A baseline model is a reference model whose performance provides a comparison point for evaluating another model.

How it is used
Uses baselines to assess whether fine-tuning improves performance on a specific task.
How deep it goes
Defines baseline alternatives and outlines evaluation steps and criteria without specifying a benchmark implementation.
Examples
Illustrates comparison with a translation model and proposes an emoji-response bot fine-tuned from GPT-3.5 Turbo.
What you can do after
Evaluate fine-tuned models against a reference using task-specific criteria and complementary quantitative and qualitative measures.

Read how Generative AI in Action explains baseline →

Effective Conversational AI

A baseline is a reference measurement of system performance taken before a change is made.

How it is used
Anchors conversational AI improvement by enabling before-and-after comparison of system performance.
How deep it goes
Defines measurement, comparison, and iteration, but stops short of prescribing one universal performance metric.
Examples
Uses chatbot accuracy, annotated utterances, intent performance, agent transfers, and process completion as measurement examples.
What you can do after
Evaluate whether a chatbot change delivered the expected improvement and diagnose gaps when results differ from expectations.

Read how Effective Conversational AI explains baseline →

Fast by Default

A reference measurement of a flow's current completion or performance against which later results are compared.

How it is used
Uses a measured starting point to compare performance changes and estimate their effect on high-value user flows.
How deep it goes
Defines flow selection, success criteria, data sources, realistic conditions, repeated measurements, and median calculation without detailing instrumentation internals.
Examples
Uses checkout, onboarding, API success, app launch, export, file open, bundle size, response time, and cold-start measurements.
What you can do after
Apply it via analytics, telemetry, monitoring dashboards, logs, and platform tools to quantify a current flow and detect later degradation.

Read how Fast by Default explains baseline →

Cloud Observability in Action

A baseline is a reference value or range representing normal system behavior against which observations are compared.

How it is used
Uses healthy behavior as a reference for alerting, anomaly detection, capacity planning, and troubleshooting.
How deep it goes
Defines values, ranges, and healthy traces, then explains manual or automated establishment without detailing anomaly-detection internals.
Examples
Uses healthy system observations, metric ranges, CloudWatch synthetic monitoring, and Jaeger happy-path and failure traces.
What you can do after
Diagnose/debug with metric ranges and Jaeger traces by detecting deviations and comparing failed services with healthy behavior.

Read how Cloud Observability in Action explains baseline →

Investing for Programmers

A benchmark is a reference measure used to compare the performance or return of an investment.

How it is used
Uses a buy-and-hold result to evaluate whether a trading strategy improves on a comparable investment.
How deep it goes
Defines a reference return, share calculation, and cost adjustment, but does not develop statistical benchmark design.
Examples
Uses fixed-$1,000 investments, NVIDIA share prices, SMA crossovers, and comparisons across companies.
What you can do after
Evaluate a trading strategy against a cost-adjusted buy-and-hold return.

Read how Investing for Programmers explains baseline →

Inside Deep Learning: Math, Algorithms, Models

How it is used
Establishes reference accuracy for comparing the impact of later network-building techniques.
How deep it goes
Defines fully connected and convolutional models and trains them for accuracy comparison.
Examples
Uses Fashion-MNIST, fully connected and CNN models, and test-accuracy curves.
What you can do after
Evaluate whether new network techniques improve accuracy against trained reference models.

Read how Inside Deep Learning: Math, Algorithms, Models explains baseline →

Build a DeepSeek Model (From Scratch)

A baseline model is a reference model trained under a simpler or comparison-only procedure without teacher guidance.

How it is used
Establishes a hard-label-only reference for comparing a student trained with teacher guidance.
How deep it goes
Compares training approaches and reports performance patterns without detailing the baseline training implementation.
Examples
Compares hard-label training, distillation at T=3, and teacher performance, including results on AIME.
What you can do after
Evaluate the effect of teacher guidance by comparing a student against a hard-label-only reference and the teacher.

Read how Build a DeepSeek Model (From Scratch) explains baseline →

AI Model Evaluation

A baseline is an existing model, method, heuristic, or reference result against which a new approach is compared.

How it is used
Provides the comparison point that turns diagnostic metrics and offline evaluation results into actionable product decisions.
How deep it goes
Defines baseline choices and comparison uses, then stops short of treating offline improvement as proof of online success.
Examples
Uses production models, previous versions, popularity recommendations, user segments, long-tail exposure, and diversity thresholds as comparison points.
What you can do after
Evaluate/understand whether a new model or policy improves meaningfully while preserving product, safety, fairness, and system constraints.

Read how AI Model Evaluation explains baseline →

Build High Value Software Systems

An initial measurement of a process or system that provides a reference for comparison with later measurements.

How it is used
Uses a pre-change measurement to determine whether a process improvement produced a result.
How deep it goes
Explains metric selection and same-method before-and-after comparison, then uses results to standardize or revert a change.
Examples
Measures PR review time, CI flakiness, manual deployment steps, and release duration as process baselines.
What you can do after
Practice a measure-change-measure cycle and use the comparison to standardize or revert a process change.

Read how Build High Value Software Systems explains baseline →

Data Storytelling with Altair and AI

A baseline is a reference value or visual reference line used to compare observed values or improvements.

How it is used
Uses reference values and lines to make trends, improvements, and decisions easier to compare.
How deep it goes
Explains baseline subtraction and threshold-line construction with runnable charting steps, stopping short of broader statistical treatment.
Examples
Demonstrates population change from 1960 and sports improvement against a 50% threshold.
What you can do after
Build it from scratch by calculating comparison differences and adding a reference rule to an Altair chart.

Read how Data Storytelling with Altair and AI explains baseline →

Human-in-the-Loop Machine Learning: Active learning and annotation for human-centered AI

A reference result or level of performance against which other results are compared.

How it is used
Adjusts annotator accuracy against expected chance performance when checking labels against ground truth.
How deep it goes
Defines three chance baselines, explains adjusted scores, and recommends choosing by task and annotator familiarity.
Examples
Compares chance baselines using four image labels with uneven frequencies and annotator accuracy scores.
What you can do after
Evaluate annotator accuracy against a task-appropriate chance baseline and flag negative adjusted scores.

Read how Human-in-the-Loop Machine Learning: Active learning and annotation for human-centered AI explains baseline →

Demand Forecasting Best Practices

A baseline model is the reference forecasting model against which forecast performance or later improvements are evaluated.

How it is used
Structures demand planning around an automated reference forecast that supports decisions, review, and improvement.
How deep it goes
Explains purpose, data choices, model alternatives, benchmarks, and review links, while deferring detailed model-development procedures.
Examples
Uses moving averages as benchmarks and contrasts limited-data time series forecasting with demand-driver-rich machine learning.
What you can do after
Evaluate forecasting approaches against a reference while balancing accuracy, bias, usefulness, workload, and planning value.

Read how Demand Forecasting Best Practices explains baseline →

Azure Security

A baseline is a reference representation of expected security or compliance requirements against which observed resource configurations can be evaluated.

How it is used
Detects anomalous identities, sign-ins, entity behavior, and resource configurations by comparing observations with expected patterns or requirements.
How deep it goes
Explains machine-learning and compliance comparisons with concrete detections, but does not detail baseline algorithms or scoring.
Examples
Uses unfamiliar locations, atypical travel, leaked credentials, anomalous file access, CIS or NIST standards, and Key Vault secrets without expiry dates.
What you can do after
Diagnose suspicious identity and entity activity with Identity Protection and Sentinel UEBA, then identify configuration misalignments through Defender for Cloud.

Read how Azure Security explains baseline →

Retrieval Augmented Generation, The Foundational Ideas

A baseline is a reference system or result against which another system's performance is compared.

How it is used
Establishes a reference for comparing retrieval quality, model changes, and production tradeoffs before adding advanced techniques.
How deep it goes
Defines comparison criteria, test-split discipline, and staged evaluation, but does not prescribe one universal baseline implementation.
Examples
Uses REALM and ORQA comparisons, unseen test splits, RAGAS measurements, and retrieval depths from k=1 through k=20.
What you can do after
Evaluate RAG configurations with comparable metrics, operational costs, unseen data, and failure-focused iteration.

Read how Retrieval Augmented Generation, The Foundational Ideas explains baseline →

Data-Driven Cybersecurity

A baseline is a reference characterization of expected system or user behavior against which deviations are evaluated.

How it is used
Establishes expected behavior and performance benchmarks for anomaly alerts, security measurement, and improvement decisions.
How deep it goes
Explains data collection, statistical comparison, thresholds, and review, while deferring deeper statistical and AI analysis.
Examples
Uses throughput logs, failed-password alerts, patching timeliness, vulnerability counts, phishing trends, and intrusion-detection rates.
What you can do after
Apply statistical comparisons and thresholds to identify deviations, evaluate controls, and refine security monitoring.

Read how Data-Driven Cybersecurity explains baseline →

Building LLM Applications with DSPy

A reference performance result against which other results are compared.

How it is used
Establishes an initial performance reference for evaluating an application and deciding whether optimization is worthwhile.
How deep it goes
Shows baseline construction and initial output inspection, then moves to evaluation without giving a numeric baseline result here.
Examples
Builds a conversation summarizer and compares its generated summary with a ground-truth summary.
What you can do after
Evaluate an initial application and use its performance as a reference when considering optimization.

Read how Building LLM Applications with DSPy explains baseline →

Also covered in

What to know first

Related concepts

Questions about baseline

What is baseline?

A baseline is a reference model, result, measurement, or level used to compare performance or assess change.

Which Manning books explain baseline?

Machine Learning System Design, Rearchitecting LLMs, Delivering Machine Learning Projects: From design to deployment, Deep Learning with R, Third Edition, Machine Learning for Tabular Data: XGBoost, Deep Learning, and AI, Effective Data Analysis, Bayesian Optimization in Action, AI Recommender Systems and 19 more.

What should I know before learning baseline?

Manning books that teach baseline usually introduce language model, metric, accuracy and classifier first.

What is the difference between baseline and benchmark?

A baseline is a reference model, result, measurement, or level used to compare performance or assess change. A benchmark is a defined, repeatable task, workload, test, dataset, or reference standard used to measure or compare systems or methods under specified conditions.

What kinds of baseline are there?

Kinds of baseline include forecast baseline, network activity baseline and uncompacted evaluation baseline.

What is baseline also called?

Baseline is also called reference point.

These guides are generated from the text of Manning books by Manning's knowledge graph and link to the sections they draw on. The full books are available on liveBook and at manning.com.