ML Basics

You won't train models — but you'll evaluate them, daily. Five concepts make that possible.

▶ Watch this reel

What you'll learn

Supervised vs unsupervised
Train/test split & overfitting
Metrics: accuracy, precision, recall, F1
scikit-learn basics
Neural network intuition

Remember this

Supervised = labeled prediction; unsupervised = structure discovery — you'll mostly evaluate, not train
Overfitting = memorized noise; sacred test set, tune on dev — same law as LLM evals
Precision = trust, recall = coverage; sklearn's fit/predict/score builds cheap baselines before LLMs

Supervised vs unsupervised

Supervised: labeled data → classification / regression.
Unsupervised: clustering, embeddings, topics.

Your role: mostly evaluate and consume models.

Train/test split & overfitting

Overfitting = memorized noise → fails on new data.
Train / dev (tuning) / test (report once, sacred).
LLM echo: benchmark contamination = same disease.

Metrics

Metric	Answers
Accuracy	overall correct — misleading with rare classes
Precision	of flagged, how many right? (trust)
Recall	of real ones, how many caught? (coverage)
F1	balance of both

Optimize by business cost: false alarm vs miss.

scikit-learn

Universal API: fit / predict / score.
Baseline first: tf-idf + logistic regression ≈ strong text classifier.
Pipeline(vectorizer, model) — no preprocessing leakage.

Neural network intuition

Layers of weighted sums + nonlinearities.
Loop: predict → loss → backprop nudges weights downhill → repeat × millions.
Transformers are this loop at trillion-token scale.

Code: Eval vocabulary + a real baseline, runnable

from sklearn.datasets import fetch_20newsgroups
from sklearn.feature_extraction.text import TfidfVectorizer
from sklearn.linear_model import LogisticRegression
from sklearn.pipeline import Pipeline
from sklearn.metrics import precision_score, recall_score, f1_score

# --- the honest split ------------------------------------------
(train, test) = (
    fetch_20newsgroups(subset="train", categories=["sci.space", "rec.sport.baseball"]),
    fetch_20newsgroups(subset="test",  categories=["sci.space", "rec.sport.baseball"]),
)

# --- baseline before any LLM ------------------------------------
pipe = Pipeline([
    ("tfidf", TfidfVectorizer()),
    ("clf",   LogisticRegression(max_iter=1000)),
])
pipe.fit(train.data, train.target)
pred = pipe.predict(test.data)

# --- the metrics that matter ------------------------------------
print(f"precision {precision_score(test.target, pred):.3f}")
print(f"recall    {recall_score(test.target, pred):.3f}")
print(f"f1        {f1_score(test.target, pred):.3f}")

# 2 categories, ~1,900 docs — trains in seconds.
# When THIS baseline fails your eval set, an LLM is justified.
# Until then, you ship this.