Fourth phase of Arc 4. The ML lingua franca.

This phase makes you fluent in the Python ML stack: numpy, pandas, scikit-learn, at the level where you stop thinking about syntax and start thinking about what to compute. Vectorization replaces for-loops. Pandas DataFrames replace dict-of-lists. scikit-learn’s API conventions (fit/transform/predict) become reflex. By phase end you’ve trained a small model end-to-end on a real dataset with proper train/test discipline.

This is also where local ML development lives. The deployment pieces (operator-managed Ray, KServe) come in Phases 37-38. Phase 34 is your dev loop.

Prerequisites

• Arc 1 Python fluency (Phase 2); Arc 3 distributed-systems understanding (Phase 21)
• A modest dataset to learn on: your own chronicle weekly logs work great for a classification or anomaly-detection task

Why this phase exists

Most engineers who say “I know ML” mean “I’ve used sklearn once.” This phase installs the operational fluency: not just calling model.fit() but reasoning about features, leakage, vectorization, memory, the bias-variance trade-off. The fluency that lets you read any ML codebase without flinching.

The pattern-first frame

Same eight steps.

1. PROBLEM

You have data. You want to predict something from it: a label, a number, a cluster. The workflow is: explore → preprocess → train → evaluate → iterate. The Python ML stack (numpy + pandas + sklearn) is the canonical toolset for this workflow at small-to-medium scale.

2. PRINCIPLES

2.1 Vectorization as the muscle

In ML, for-loops are slow. Vectorized operations (numpy arrays, pandas Series) run in C, 100-1000× faster. The mental shift is “think in arrays, not elements.”

→ Pattern: vectorization-as-pattern

Investigate:

• Why is numpy fast? (Hint: contiguous memory, SIMD, no Python interpreter overhead per element.)
• Walk a vectorized operation step-by-step in memory.
• When is vectorization the wrong tool? (Hint: irregular control flow.)

2.2 Train / test / validation split

You can’t evaluate a model on the data you trained it on (it memorizes). Split: train (fit), validation (tune), test (final report). Never let the test set touch hyperparameter tuning.

→ Pattern: train-test-split

Investigate:

• What’s data leakage, and what does it look like?
• When is a simple train/test split insufficient (time-series, grouped data)?
• Why is the test set held out forever until the final report?

2.3 Cross-validation

K-fold cross-validation gives a more reliable performance estimate by averaging across multiple splits. Stratified, group, time-series CV handle structure in the data.

→ Pattern: cross-validation

Investigate:

• Walk 5-fold CV. What’s averaged?
• When does CV underestimate generalization error (data leakage across folds)?
• How does GroupKFold differ from KFold, and when does it matter?

2.4 Feature engineering and scaling

Many models assume features on similar scales. Standardization (z-score), normalization (min-max), and log transforms each have a use case.

Investigate:

• Why does StandardScaler matter for SVM but not for decision trees?
• What’s leakage in scaling, and how does sklearn’s Pipeline prevent it?
• When are categorical encodings the bottleneck (one-hot, target, ordinal)?

2.5 Bias-variance trade-off

A model can underfit (high bias, too simple) or overfit (high variance, too complex). The sweet spot depends on data size, complexity, regularization.

Investigate:

• Walk learning curves: training error + validation error vs training set size.
• What’s regularization, and how does L1 vs L2 differ?
• Why is “deep learning has high variance” the operational reality for small datasets?

2.6 The sklearn API as design pattern

fit(X, y), transform(X), predict(X), score(X, y). Consistent across every estimator. Pipelines chain transformers + estimators. This is the API that defined the modern ML ecosystem.

Investigate:

• Why is the Pipeline object the right way to do preprocessing + modeling?
• What’s fit_transform vs fit then transform?
• How does this API translate to PyTorch / TensorFlow (it doesn’t, exactly)?

3. TRADE-OFFS

4. TOOLS (as of 2026-06)

• numpy, pandas, scikit-learn
• polars as fast pandas alternative
• jupyterlab for exploratory dev
• uv for env management (consistent with Arc 1 Python tooling)

Reading

• “Python for Data Analysis” (Wes McKinney, 3rd ed.): pandas canonical
• “Hands-On Machine Learning” (Géron, 3rd ed.): sklearn-flavored ML intro
• scikit-learn user guide (free, comprehensive)
• “An Introduction to Statistical Learning” (James et al., 2nd ed.): free online

5. MASTERY: One model end-to-end on real data

[ ] Pick a real dataset (your chronicle logs, public dataset, your day-job data with permission)
[ ] Exploratory data analysis: distributions, missing values, correlations
[ ] Vectorize all your processing: no for-loops
[ ] Split: train / validation / test
[ ] Build a Pipeline with preprocessing + estimator
[ ] Cross-validate; report mean + std
[ ] Tune hyperparameters (GridSearchCV or RandomizedSearchCV) on validation
[ ] Evaluate on held-out test set: once, at the end
[ ] Write a model card documenting: problem, data, model, performance, limitations
[ ] Profile your training loop: where is time spent?

6. COMPARE: Polars or DuckDB

Reimplement one of your pandas pipelines in Polars (Rust-backed, fast) or DuckDB (SQL over DataFrames). Reflect on speed + ergonomics.

400-word reflection.

7. OPERATE

• Notebook discipline: one notebook per experiment; commit notebooks to a notebooks/ repo; export results as markdown reports
• 1-2 runbooks (e.g., “Reproduce a notebook result”; “Diagnose data leakage”)
• Weekly log

8. CONTRIBUTE

• scikit-learn: docs, examples
• pandas: documentation
• A small public Kaggle-style competition for practice (your portfolio doesn’t need to win; the practice is the point)

What ships from this phase

• First trained model documented with model card
• Notebooks repo on GitHub (public OK; sensitive data scrubbed)
• Vectorization muscle: you’ll read ML code without flinching

Validation criteria

[ ] End-to-end model on real data with proper train/test discipline
[ ] Pipeline-based preprocessing (no leakage)
[ ] Cross-validation done properly
[ ] Model card written
[ ] All 10 operational depth checks
[ ] Polars/DuckDB compare reflection (400 words)
[ ] 1-2 runbooks
[ ] Pattern entries:
    - vectorization-as-pattern → OUTLINE
    - train-test-split → OUTLINE
    - cross-validation → OUTLINE
[ ] Exit Test passed

Exit Test

Time: 2 hours.

Part 1: Build (75 min)

Given a new dataset (provided), build a complete pipeline: preprocess + model + evaluate. Properly cross-validated. Model card written.

Part 2: Diagnose (30 min)

A leakage scenario: model performs great on validation, terrible on a separate test set. Identify the leakage source.

Part 3: Articulate (15 min)

~400 words: “When would you reach for pandas vs Polars vs DuckDB for a 10GB CSV processing task?”

Anti-patterns

Patterns touched this phase

• vectorization-as-pattern: OUTLINE
• train-test-split: OUTLINE
• cross-validation: OUTLINE

→ Next: Phase 35: Classical ML Engineering