1. LandOffer
  2. Blog
  3. All articles
  4. Machine Learning Portfolio for Applications: Document Evaluation, Limits, and Reproducibility

Job search

Machine Learning Portfolio for Applications: Document Evaluation, Limits, and Reproducibility

Build an inspectable machine learning portfolio with dataset provenance, a fixed split, a baseline, recorded errors and a reproducible local model card.

9 min readLandOffer team

A machine learning portfolio should make the experiment inspectable: identify the task and dataset, preserve the train/test split, compare against a baseline, report the metric and errors, and supply the commands needed to reproduce the result. A model card can organize this evidence. Keep offline evaluation separate from deployment, operating reliability and business outcomes, and select the sample according to the role's actual responsibilities.

LandOffer publishes this guide and offers job-search tools. We ran an original local experiment using scikit-learn's public Wine dataset and retained its results, split indices and model card. Alaric, the candidate used to discuss the packet, is fictional. The reported measurements are actual local observations, not a hiring study or production-model result. Begin with recent roles on LandOffer and identify which modeling responsibility your sample needs to demonstrate.

An editorial experiment bench keeps public data, a fixed training and test split, a baseline and a model card connected through an inspectable local record.

Select evidence for the responsibility you want to show

A role focused on modeling may ask for problem formulation, evaluation and error analysis. A role focused on operating models may also require packaging, deployment, monitoring and recovery. An offline classifier can demonstrate part of the first responsibility; it cannot establish the second merely because its score is high.

Read the posting for the work, audience and required environment. Decide which project artifact answers each substantial requirement. A notebook can contain the experiment, a model card can summarize its purpose and limitations, and a script can reproduce the key measurements. These artifacts are useful when they expose the method, result and contribution.

Alaric's fictional target is an entry-level role that requests an explainable classification project and evaluation discussion. His completed packet leads with a small offline experiment. If another posting requires independent production operation, that requirement remains a gap. He does not relabel the experiment as a deployed service to make the portfolio fit.

The machine learning engineer resume guide explains the separate operating-responsibility question. Use your portfolio to preserve the underlying evidence, then write only the supported contribution in the resume. A concise public description must remain true even if the reviewer never opens the project.

Describe the dataset before reporting a score

The UCI Wine dataset concerns chemical measurements from wines derived from three cultivars. It has 178 instances and 13 features. This is a cultivar-classification exercise, not a wine-quality predictor. UCI identifies it as a relatively straightforward classic classification dataset, which limits the significance of a strong result.

We used the numeric copy loaded by scikit-learn's load_wine. Its class labels are class_0, class_1 and class_2, with 59, 71 and 48 rows respectively. The packet retains feature names, the packaged description and a hash of the numeric inputs and targets. This record identifies the actual experiment's inputs.

Keep dataset authorship separate from your work. Alaric analyzed the supplied data; he did not collect it or define the original classification task. The current UCI page supplies a citation for Aeberhard and Forina and lists a CC BY 4.0 license. The installed package's descriptive provenance fields differ from the current UCI citation; we retained the package text and used the current UCI citation in the model card rather than silently combining the two.

A dataset note should also explain what is unknown. This small classic sample does not supply evidence about new laboratories, future production inputs or a commercial population. A random holdout within it cannot establish performance in those settings. Do not invent a contemporary use case and present the dataset as representative of that use without a separate justification.

Preserve the split and baseline before inspecting results

Before execution, the project brief specified a 30 percent stratified holdout, random seed 19, a most-frequent-class baseline and a decision tree limited to depth two. Those choices were recorded before the test results were examined. The original goal was to compare a simple model with a simple baseline and document the comparison, not to find the highest score.

The resulting split contains 124 training rows and 54 test rows, with no overlapping indices. The test class counts are 18, 21 and 15. We retained the actual indices so the split can be inspected. scikit-learn's split documentation explains the test proportion, randomness and stratification parameters; a fixed seed makes this particular split repeatable, not representative of every possible sample.

Recorded choice Actual exercise What it establishes
Test boundary 54 retained holdout rows The population used for this evaluation
Training boundary 124 rows, no index overlap Which records were available during fitting
Baseline Always predict the most frequent training class A simple reference for this task
Candidate model Decision tree, maximum depth two A predeclared, inspectable model choice

Both estimators were fitted only on training rows. There was no parameter search or feature selection informed by the holdout. The tree exercise used no fitted preprocessing transformation. If you add preprocessing to a different project, fit it within the training boundary and apply the established transformation to the test inputs.

scikit-learn's common-pitfalls guide explains why information leaking from test data can make performance estimates overly optimistic. Apply that principle to the entire experiment, including feature selection and tuning. If you change the model after inspecting this holdout, label the subsequent result accordingly; the same test set is no longer an untouched comparison for that choice.

Report a comparison with the denominator visible

The most-frequent baseline correctly classified 21 of 54 test rows, for accuracy 0.388889. The depth-two tree correctly classified 48 of 54, for accuracy 0.888889. Their macro F1 values were 0.186667 and 0.889703. These are actual local results from the recorded split and configuration.

Estimator Correct test predictions Accuracy Macro F1
Most-frequent baseline 21 of 54 0.388889 0.186667
Depth-two tree 48 of 54 0.888889 0.889703

Accuracy reports the share of correct predictions in this test set. Macro F1 averages the class-level F1 scores, giving each class equal weight in that average. Keeping both beside class counts helps the reader inspect how the comparison was summarized. Both metrics describe the retained holdout; neither measures deployment readiness.

The packet retains the full per-class report, confusion matrix and predictions, not only the headline values. A reviewer can follow a number back to individual held-out records and see the baseline's behavior. Report enough precision to reproduce the calculation, but avoid treating six decimal places as evidence of broad certainty.

Do not compare this score with a tutorial's score unless the datasets, splits, preprocessing and evaluation protocols are comparable. A better number can come from an easier test population rather than a better model. For this portfolio, the meaningful comparison is the recorded baseline and tree evaluated on the same retained rows.

Use the errors to define a failure scope

The tree made six errors. Three test rows from class_1 and three from class_2 were predicted as class_0. All 18 held-out class_0 rows were classified correctly in this run. The actual confusion matrix is shown below; rows are true classes and columns are predicted classes.

True class Predicted class_0 Predicted class_1 Predicted class_2
class_0 18 0 0
class_1 3 18 0
class_2 3 0 12

The saved prediction table identifies the misclassified dataset rows: 123, 149, 83, 151, 66 and 150, using zero-based indices in the packaged copy. The exported tree rules provide another inspectable artifact. They let a reviewer discuss how this shallow tree divided the training examples and where the retained test errors appeared.

An editorial review desk keeps marked error records beside a confusion grid and a limitations note, separating observed failures from untested deployment claims.

Do not turn the error pattern into a causal explanation without evidence. The record shows which labels were confused. It does not prove that the source measurements were wrong, that the labels should be changed or that one chemical feature caused a real-world failure. The model's branch rules are behavior of this fitted estimator, not scientific conclusions about wine production.

Alaric's completed decision is to present the exercise as a bounded baseline comparison with visible errors. He would need a new, appropriate evaluation plan before making claims about new sources or commercial use. He does not tune on these six errors and then describe a new result as the original holdout evaluation. A follow-up experiment would require a separate record and an honest account of what informed it.

A completed model card and contribution map

Hugging Face's model-card guidance describes a Markdown document containing model information, intended uses and limitations, training details, datasets and evaluation results. On the Hub, a repository README can render as the card with additional metadata. That documents a format and platform capability; it does not prove that every recruiter requires a Hub profile.

Our original local card includes the task, dataset citation, predeclared split and estimators, actual metrics, six observed errors and reproduction information. Its intended use is inspection of a small offline portfolio experiment. The limits explicitly exclude wine quality, commercial validation, deployment, service reliability and decisions about people.

Here is a completed excerpt:

Original offline experiment. Used the packaged UCI Wine data for a three-class classification comparison. Fitted a most-frequent baseline and a depth-two tree on 124 training rows; evaluated both on the retained 54-row stratified holdout. Recorded predictions, per-class results and six tree errors. One local split supports this comparison; new-population performance and production operation were not evaluated.

Contribution Supported record Boundary
Experiment configuration Prewriting brief and script Original local setup, not a new learning algorithm
Evaluation and error account Split indices, predictions and result JSON This dataset and holdout only
Documentation Authored card and readable export No actual reviewer feedback or user study
Data and estimator implementation UCI source and scikit-learn Credit the source authors and library

This map lets Alaric explain what he did without taking credit for the dataset or implementation of the estimator. For team projects, extend the same map to collaborators. The personal-project contribution guide develops that ownership boundary for resumes.

Make the packet reproducible and readable

The experiment ran in an isolated local environment with Python 3.12.14, scikit-learn 1.9.1 and NumPy 2.5.4. We retained the environment requirements, a standalone script, source-data hash, split indices and result files. Running the same script again reproduced the split, accuracy, macro F1 and confusion matrices. That repeat confirms the recorded configuration locally; it is not an independent validation population.

A useful README supplies the startup commands, dependencies, expected files and approximate task scope. Keep relative file locations clear and remove personal credentials or private paths from a shareable copy. If a reviewer cannot run the environment, provide a readable card and result table that preserve the method and limits. Include the method and limits beside the score so readers can assess it.

We exported the original card to HTML and opened it locally. No Hugging Face account, repository upload or hosted prediction endpoint was used. Before an actual application, verify the final external destination separately and follow the requested sample format. Public hosting is another workflow, with its own access and data-sharing decisions.

Choose the packet that covers the target responsibility, then describe the actual experiment and its environment. Keep production gaps explicit and give the reviewer a route to the evidence. Identify the required responsibility in recent roles on LandOffer, then select the packet that shows the work you actually completed.

Sources and Further Reading