MLflow 101: Local Experiment Tracking for Beginners
This blog is a beginner-friendly walkthrough of local experiment tracking with MLflow. If you’ve ever lost a “good” result because you forgot which settings you used, MLflow solves the day-one problem in many student projects: what did I run and how well did it do? You can start on a laptop with no cloud setup and still keep a reproducible record of parameters, metrics, and artifacts.
Why this helps in practice. From what I’ve seen in classes and hackathons, people usually try one of three things: (1) screenshots and notebook cells (fast, but impossible to compare later), (2) spreadsheets (organized, but manual and easy to break), or (3) full platforms like Weights & Biases (great for teams, but heavy when you’re just learning). MLflow hits a clean middle ground for early projects: a tiny API to log runs and a UI to compare them. If your work scales up, the same code can point to a remote tracking server or be paired with data versioning tools.
Prerequisites
Below is a minimal example that trains a baseline LogisticRegression and logs a few metrics and artifacts. The goal isn’t to chase state-of-the-art performance; it’s to show a small, reusable pattern you can drop into your own repos.
# train.py — minimal MLflow example
import argparse
import mlflow, mlflow.sklearn
from sklearn.datasets import load_breast_cancer
from sklearn.model_selection import train_test_split
from sklearn.linear_model import LogisticRegression
from sklearn.metrics import accuracy_score, roc_auc_score
import pandas as pd
parser = argparse.ArgumentParser()
parser.add_argument("--C", type=float, default=1.0, help="Inverse regularization strength")
args = parser.parse_args()
# Data
X, y = load_breast_cancer(return_X_y=True, as_frame=True)
Xtr, Xte, ytr, yte = train_test_split(X, y, test_size=0.2, random_state=42)
# Experiment + run
mlflow.set_experiment("mlflow-101")
with mlflow.start_run():
model = LogisticRegression(max_iter=200, C=args.C, solver="lbfgs")
model.fit(Xtr, ytr)
yhat = model.predict(Xte)
proba = model.predict_proba(Xte)[:, 1]
acc = accuracy_score(yte, yhat)
auc = roc_auc_score(yte, proba)
mlflow.log_param("C", args.C)
mlflow.log_metric("accuracy", acc)
mlflow.log_metric("roc_auc", auc)
# Save model + prediction artifacts
mlflow.sklearn.log_model(model, "model")
pd.DataFrame({"y_true": yte, "y_score": proba}).to_csv("preds.csv", index=False)
mlflow.log_artifact("preds.csv")
print("done")Run a few experiments with different regularization strengths and compare them in the local UI:
# 1) Create env and install
python -m venv .venv
# macOS/Linux:
source .venv/bin/activate
# Windows (PowerShell):
# .venv\Scripts\Activate.ps1
pip install mlflow scikit-learn pandas
# 2) Run experiments
python train.py --C 0.1
python train.py --C 1.0
python train.py --C 10
# 3) Open the Tracking UI from the folder that contains the `mlruns/` directory
mlflow ui # visit http://127.0.0.1:5000Questions?
If you have lingering questions about this resource, please post to the Nexus Q&A on GitHub.